{
  "schema_version": 1,
  "updated": "2026-11-14",
  "license": "MIT",
  "source": "https://whatmodelscanirun.com",
  "docs": "https://whatmodelscanirun.com/data/",
  "count": 214,
  "models": [
    {
      "id": "arete-27b-q4",
      "name": "Arete 27B",
      "url": "https://whatmodelscanirun.com/model/arete-27b",
      "quantization": "Q4_K_M",
      "params_b": 27.36,
      "weight_gb": 16.02,
      "gguf_repo": "bartowski/NousResearch_Arete-27B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.061,
      "max_context_k": 256,
      "layers": 64,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": null,
      "swe_bench_score": null,
      "notes": "Nous Research's supervised fine-tune of Qwen3.8-27B that removes CCP-aligned framing and refusals on politically sensitive topics; architecture unchanged. 64 layers: 16 full-attention (4 KV heads, head_dim 256) and 48 linear-attention, so only the full-attention layers grow the KV cache. 262K context. Vision encoder kept, frozen, from Qwen3.8 27B. Nous reports no meaningful regression on general and agentic benchmarks but publishes no MMLU-Pro or SWE-bench Verified. Weight excludes the separate mmproj file. Apache 2.0. GGUF Q4_K_M file size."
    },
    {
      "id": "arete-27b-q8",
      "name": "Arete 27B",
      "url": "https://whatmodelscanirun.com/model/arete-27b",
      "quantization": "Q8_0",
      "params_b": 27.36,
      "weight_gb": 26.7,
      "gguf_repo": "bartowski/NousResearch_Arete-27B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.061,
      "max_context_k": 256,
      "layers": 64,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": null,
      "swe_bench_score": null,
      "notes": "Nous Research's supervised fine-tune of Qwen3.8-27B that removes CCP-aligned framing and refusals on politically sensitive topics; architecture unchanged. 64 layers: 16 full-attention (4 KV heads, head_dim 256) and 48 linear-attention, so only the full-attention layers grow the KV cache. 262K context. Vision encoder kept, frozen, from Qwen3.8 27B. Nous reports no meaningful regression on general and agentic benchmarks but publishes no MMLU-Pro or SWE-bench Verified. Weight excludes the separate mmproj file. Apache 2.0. GGUF Q8_0 file size."
    },
    {
      "id": "beam-1.0-501b-a23b-q4",
      "name": "Beam 1.0 501B A23B",
      "url": "https://whatmodelscanirun.com/model/beam-1-0-501b-a23b",
      "quantization": "Q4_K_M",
      "params_b": 501.27,
      "weight_gb": 292.93,
      "gguf_repo": "bartowski/Beam-1.0-501B-A23B-GGUF",
      "gguf_shards": 9,
      "active_weight_gb": 13.67,
      "kv_per_1k_gb": 0.034,
      "max_context_k": 1024,
      "layers": 52,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": 81.6,
      "notes": "Reflection AI's first open model: sparse MoE, 501.3B total, 23.4B active (256 experts, 8 per token). 52 layers (3 dense + 49 MoE): 9 full-attention gated-GQA layers (8 KV heads, head_dim 128) and 43 sliding-window (512) layers, so only the full-attention layers grow the KV cache. 1M context. Text only. SWE-bench Verified 81.6 (BF16), SWE-bench Multilingual 78.3. MMLU-Pro not published. Apache 2.0. GGUF Q4_K_M file size."
    },
    {
      "id": "beam-1.0-501b-a23b-q8",
      "name": "Beam 1.0 501B A23B",
      "url": "https://whatmodelscanirun.com/model/beam-1-0-501b-a23b",
      "quantization": "Q8_0",
      "params_b": 501.27,
      "weight_gb": 496.2,
      "gguf_repo": "bartowski/Beam-1.0-501B-A23B-GGUF",
      "gguf_shards": 14,
      "active_weight_gb": 23.16,
      "kv_per_1k_gb": 0.034,
      "max_context_k": 1024,
      "layers": 52,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": 81.6,
      "notes": "Reflection AI's first open model: sparse MoE, 501.3B total, 23.4B active (256 experts, 8 per token). 52 layers (3 dense + 49 MoE): 9 full-attention gated-GQA layers (8 KV heads, head_dim 128) and 43 sliding-window (512) layers, so only the full-attention layers grow the KV cache. 1M context. Text only. SWE-bench Verified 81.6 (BF16), SWE-bench Multilingual 78.3. MMLU-Pro not published. Apache 2.0. GGUF Q8_0 file size."
    },
    {
      "id": "codestral-22b-q4",
      "name": "Codestral 22B",
      "url": "https://whatmodelscanirun.com/model/codestral-22b",
      "quantization": "Q4_K_M",
      "params_b": 22.2,
      "weight_gb": 12.42,
      "gguf_repo": "bartowski/Codestral-22B-v0.1-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.214,
      "max_context_k": 32,
      "layers": 56,
      "features": [],
      "mmlu_score": 62,
      "swe_bench_score": null,
      "notes": "Code model. Mistral architecture. 56 layers, 8 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "codestral-22b-q8",
      "name": "Codestral 22B",
      "url": "https://whatmodelscanirun.com/model/codestral-22b",
      "quantization": "Q8_0",
      "params_b": 22.2,
      "weight_gb": 22.02,
      "gguf_repo": "bartowski/Codestral-22B-v0.1-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.214,
      "max_context_k": 32,
      "layers": 56,
      "features": [],
      "mmlu_score": 62,
      "swe_bench_score": null,
      "notes": "Code model. Mistral architecture. 56 layers, 8 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "deepseek-coder-6.7b-q4",
      "name": "DeepSeek Coder 6.7B",
      "url": "https://whatmodelscanirun.com/model/deepseek-coder-6-7b",
      "quantization": "Q4_K_M",
      "params_b": 6.74,
      "weight_gb": 3.8,
      "gguf_repo": "TheBloke/deepseek-coder-6.7B-instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.488,
      "max_context_k": 16,
      "layers": 32,
      "features": [],
      "mmlu_score": 50,
      "swe_bench_score": null,
      "notes": "Code model. Llama architecture. 32 layers, 32 KV heads (full MHA), head_dim 128. 16K context. GGUF Q4_K_M file size."
    },
    {
      "id": "deepseek-coder-6.7b-q8",
      "name": "DeepSeek Coder 6.7B",
      "url": "https://whatmodelscanirun.com/model/deepseek-coder-6-7b",
      "quantization": "Q8_0",
      "params_b": 6.74,
      "weight_gb": 6.67,
      "gguf_repo": "TheBloke/deepseek-coder-6.7B-instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.488,
      "max_context_k": 16,
      "layers": 32,
      "features": [],
      "mmlu_score": 50,
      "swe_bench_score": null,
      "notes": "Code model. Llama architecture. 32 layers, 32 KV heads (full MHA), head_dim 128. 16K context. GGUF Q8_0 file size."
    },
    {
      "id": "deepseek-r1-0528-q4",
      "name": "DeepSeek R1 0528",
      "url": "https://whatmodelscanirun.com/model/deepseek-r1-0528",
      "quantization": "Q4_K_M",
      "params_b": 671,
      "weight_gb": 381.12,
      "gguf_repo": "bartowski/deepseek-ai_DeepSeek-R1-0528-GGUF",
      "gguf_shards": 11,
      "active_weight_gb": 21.02,
      "kv_per_1k_gb": 0.065,
      "max_context_k": 160,
      "layers": 61,
      "features": [
        "reasoning"
      ],
      "mmlu_score": 90.8,
      "swe_bench_score": 57.6,
      "notes": "671B MoE (37B active), MLA KV compression; split GGUF ~381 GiB Active params: 37B (DeepSeek-R1 model card)."
    },
    {
      "id": "deepseek-r1-0528-q8",
      "name": "DeepSeek R1 0528",
      "url": "https://whatmodelscanirun.com/model/deepseek-r1-0528",
      "quantization": "Q8_0",
      "params_b": 671,
      "weight_gb": 664.3,
      "gguf_repo": "unsloth/DeepSeek-R1-0528-GGUF",
      "gguf_shards": 15,
      "active_weight_gb": 36.63,
      "kv_per_1k_gb": 0.065,
      "max_context_k": 160,
      "layers": 61,
      "features": [
        "reasoning"
      ],
      "mmlu_score": 90.8,
      "swe_bench_score": 57.6,
      "notes": "671B MoE (37B active), MLA KV compression; split GGUF ~664 GiB Active params: 37B (DeepSeek-R1 model card)."
    },
    {
      "id": "deepseek-r1-distill-llama-70b-q4",
      "name": "DeepSeek R1 Distill Llama 70B",
      "url": "https://whatmodelscanirun.com/model/deepseek-r1-distill-llama-70b",
      "quantization": "Q4_K_M",
      "params_b": 70.6,
      "weight_gb": 39.6,
      "gguf_repo": "unsloth/DeepSeek-R1-Distill-Llama-70B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.305,
      "max_context_k": 128,
      "layers": 80,
      "features": [
        "reasoning"
      ],
      "mmlu_score": 84,
      "swe_bench_score": 27.2,
      "notes": "Reasoning model (distilled from R1). Based on Llama 3.3 70B. 80 layers, 8 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "deepseek-r1-distill-llama-70b-q8",
      "name": "DeepSeek R1 Distill Llama 70B",
      "url": "https://whatmodelscanirun.com/model/deepseek-r1-distill-llama-70b",
      "quantization": "Q8_0",
      "params_b": 70.6,
      "weight_gb": 69.83,
      "gguf_repo": "unsloth/DeepSeek-R1-Distill-Llama-70B-GGUF",
      "gguf_shards": 2,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.305,
      "max_context_k": 128,
      "layers": 80,
      "features": [
        "reasoning"
      ],
      "mmlu_score": 84,
      "swe_bench_score": 27.2,
      "notes": "Reasoning model (distilled from R1). Based on Llama 3.3 70B. 80 layers, 8 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "deepseek-r1-distill-qwen-14b-q4",
      "name": "DeepSeek R1 Distill Qwen 14B",
      "url": "https://whatmodelscanirun.com/model/deepseek-r1-distill-qwen-14b",
      "quantization": "Q4_K_M",
      "params_b": 14.7,
      "weight_gb": 8.37,
      "gguf_repo": "unsloth/DeepSeek-R1-Distill-Qwen-14B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.183,
      "max_context_k": 128,
      "layers": 48,
      "features": [
        "reasoning"
      ],
      "mmlu_score": 72,
      "swe_bench_score": 20.4,
      "notes": "Reasoning model (distilled from R1). Based on Qwen 2.5 14B. 48 layers, 8 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "deepseek-r1-distill-qwen-14b-q8",
      "name": "DeepSeek R1 Distill Qwen 14B",
      "url": "https://whatmodelscanirun.com/model/deepseek-r1-distill-qwen-14b",
      "quantization": "Q8_0",
      "params_b": 14.7,
      "weight_gb": 14.62,
      "gguf_repo": "unsloth/DeepSeek-R1-Distill-Qwen-14B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.183,
      "max_context_k": 128,
      "layers": 48,
      "features": [
        "reasoning"
      ],
      "mmlu_score": 72,
      "swe_bench_score": 20.4,
      "notes": "Reasoning model (distilled from R1). Based on Qwen 2.5 14B. 48 layers, 8 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "deepseek-r1-distill-qwen-32b-q4",
      "name": "DeepSeek R1 Distill Qwen 32B",
      "url": "https://whatmodelscanirun.com/model/deepseek-r1-distill-qwen-32b",
      "quantization": "Q4_K_M",
      "params_b": 32.5,
      "weight_gb": 18.49,
      "gguf_repo": "unsloth/DeepSeek-R1-Distill-Qwen-32B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.244,
      "max_context_k": 128,
      "layers": 64,
      "features": [
        "reasoning"
      ],
      "mmlu_score": 80,
      "swe_bench_score": 30.6,
      "notes": "Reasoning model (distilled from R1). Based on Qwen 2.5 32B. 64 layers, 8 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "deepseek-r1-distill-qwen-32b-q8",
      "name": "DeepSeek R1 Distill Qwen 32B",
      "url": "https://whatmodelscanirun.com/model/deepseek-r1-distill-qwen-32b",
      "quantization": "Q8_0",
      "params_b": 32.5,
      "weight_gb": 32.43,
      "gguf_repo": "unsloth/DeepSeek-R1-Distill-Qwen-32B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.244,
      "max_context_k": 128,
      "layers": 64,
      "features": [
        "reasoning"
      ],
      "mmlu_score": 80,
      "swe_bench_score": 30.6,
      "notes": "Reasoning model (distilled from R1). Based on Qwen 2.5 32B. 64 layers, 8 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "deepseek-r1-distill-qwen-7b-q4",
      "name": "DeepSeek R1 Distill Qwen 7B",
      "url": "https://whatmodelscanirun.com/model/deepseek-r1-distill-qwen-7b",
      "quantization": "Q4_K_M",
      "params_b": 7.62,
      "weight_gb": 4.36,
      "gguf_repo": "unsloth/DeepSeek-R1-Distill-Qwen-7B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.053,
      "max_context_k": 128,
      "layers": 28,
      "features": [
        "reasoning"
      ],
      "mmlu_score": 58,
      "swe_bench_score": 12.2,
      "notes": "Reasoning model (distilled from R1). Based on Qwen 2.5 Math 7B. 28 layers, 4 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "deepseek-r1-distill-qwen-7b-q8",
      "name": "DeepSeek R1 Distill Qwen 7B",
      "url": "https://whatmodelscanirun.com/model/deepseek-r1-distill-qwen-7b",
      "quantization": "Q8_0",
      "params_b": 7.62,
      "weight_gb": 7.54,
      "gguf_repo": "unsloth/DeepSeek-R1-Distill-Qwen-7B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.053,
      "max_context_k": 128,
      "layers": 28,
      "features": [
        "reasoning"
      ],
      "mmlu_score": 58,
      "swe_bench_score": 12.2,
      "notes": "Reasoning model (distilled from R1). Based on Qwen 2.5 Math 7B. 28 layers, 4 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "deepseek-v4-flash-q4",
      "name": "DeepSeek V4 Flash",
      "url": "https://whatmodelscanirun.com/model/deepseek-v4-flash",
      "quantization": "Q4_K_M",
      "params_b": 284,
      "weight_gb": 161,
      "gguf_repo": null,
      "gguf_shards": null,
      "active_weight_gb": 7.37,
      "kv_per_1k_gb": 0.082,
      "max_context_k": 1024,
      "layers": 43,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 86.4,
      "swe_bench_score": 79,
      "notes": "43 layers, MQA (1 KV head), head_dim 512; MoE 256 experts 6 active Active params: 13B (DeepSeek-V4 model card)."
    },
    {
      "id": "deepseek-v4-flash-q8",
      "name": "DeepSeek V4 Flash",
      "url": "https://whatmodelscanirun.com/model/deepseek-v4-flash",
      "quantization": "Q8_0",
      "params_b": 284,
      "weight_gb": 282,
      "gguf_repo": null,
      "gguf_shards": null,
      "active_weight_gb": 12.91,
      "kv_per_1k_gb": 0.082,
      "max_context_k": 1024,
      "layers": 43,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 86.4,
      "swe_bench_score": 79,
      "notes": "43 layers, MQA (1 KV head), head_dim 512; MoE 256 experts 6 active Active params: 13B (DeepSeek-V4 model card)."
    },
    {
      "id": "deepseek-v4-pro-q4",
      "name": "DeepSeek V4 Pro",
      "url": "https://whatmodelscanirun.com/model/deepseek-v4-pro",
      "quantization": "Q4_K_M",
      "params_b": 1600,
      "weight_gb": 900,
      "gguf_repo": null,
      "gguf_shards": null,
      "active_weight_gb": 27.56,
      "kv_per_1k_gb": 0.065,
      "max_context_k": 1024,
      "layers": 61,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 87.5,
      "swe_bench_score": 80.6,
      "notes": "61 layers, MLA KV compression (rank 512); MoE 384 experts 6 active Active params: 49B (DeepSeek-V4 model card)."
    },
    {
      "id": "deepseek-v4-pro-q8",
      "name": "DeepSeek V4 Pro",
      "url": "https://whatmodelscanirun.com/model/deepseek-v4-pro",
      "quantization": "Q8_0",
      "params_b": 1600,
      "weight_gb": 1670,
      "gguf_repo": null,
      "gguf_shards": null,
      "active_weight_gb": 51.14,
      "kv_per_1k_gb": 0.065,
      "max_context_k": 1024,
      "layers": 61,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 87.5,
      "swe_bench_score": 80.6,
      "notes": "61 layers, MLA KV compression (rank 512); MoE 384 experts 6 active Active params: 49B (DeepSeek-V4 model card)."
    },
    {
      "id": "devstral-small-24b-q4",
      "name": "Devstral Small 24B",
      "url": "https://whatmodelscanirun.com/model/devstral-small-24b",
      "quantization": "Q4_K_M",
      "params_b": 23.6,
      "weight_gb": 13.35,
      "gguf_repo": "unsloth/Devstral-Small-2505-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.153,
      "max_context_k": 128,
      "layers": 40,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 72,
      "swe_bench_score": 46.8,
      "notes": "Code model. Fine-tuned from Mistral Small 3.1 (vision encoder removed). 40 layers, 8 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "devstral-small-24b-q8",
      "name": "Devstral Small 24B",
      "url": "https://whatmodelscanirun.com/model/devstral-small-24b",
      "quantization": "Q8_0",
      "params_b": 23.6,
      "weight_gb": 23.33,
      "gguf_repo": "unsloth/Devstral-Small-2505-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.153,
      "max_context_k": 128,
      "layers": 40,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 72,
      "swe_bench_score": 46.8,
      "notes": "Code model. Fine-tuned from Mistral Small 3.1 (vision encoder removed). 40 layers, 8 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "flint-40b-a2.5b-q4",
      "name": "Flint 40B A2.5B",
      "url": "https://whatmodelscanirun.com/model/flint-40b-a2-5b",
      "quantization": "Q4_K_M",
      "params_b": 39.79,
      "weight_gb": 23.61,
      "gguf_repo": "bartowski/RecursiveSI_flint-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 1.48,
      "kv_per_1k_gb": 0.027,
      "max_context_k": 256,
      "layers": 49,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": 74.2,
      "notes": "Recursive's coding-agent MoE: 39.8B total, ~2.5B active (256 experts, 6 per token, plus shared expert). 49 layers: 14 full-attention (4 KV heads, head_dim 128, no positional encoding) and 35 sliding-window (513 tokens), so only the full-attention layers grow the KV cache. 262K context. SWE-bench Verified 74.2 (mini-swe-agent), Terminal-Bench 2.1 54.9. MMLU-Pro not published. Apache 2.0. GGUF Q4_K_M file size."
    },
    {
      "id": "flint-40b-a2.5b-q8",
      "name": "Flint 40B A2.5B",
      "url": "https://whatmodelscanirun.com/model/flint-40b-a2-5b",
      "quantization": "Q8_0",
      "params_b": 39.79,
      "weight_gb": 39.44,
      "gguf_repo": "bartowski/RecursiveSI_flint-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 2.48,
      "kv_per_1k_gb": 0.027,
      "max_context_k": 256,
      "layers": 49,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": 74.2,
      "notes": "Recursive's coding-agent MoE: 39.8B total, ~2.5B active (256 experts, 6 per token, plus shared expert). 49 layers: 14 full-attention (4 KV heads, head_dim 128, no positional encoding) and 35 sliding-window (513 tokens), so only the full-attention layers grow the KV cache. 262K context. SWE-bench Verified 74.2 (mini-swe-agent), Terminal-Bench 2.1 54.9. MMLU-Pro not published. Apache 2.0. GGUF Q8_0 file size."
    },
    {
      "id": "gemma-2-27b-q4",
      "name": "Gemma 2 27B",
      "url": "https://whatmodelscanirun.com/model/gemma-2-27b",
      "quantization": "Q4_K_M",
      "params_b": 27.2,
      "weight_gb": 15.5,
      "gguf_repo": "bartowski/gemma-2-27b-it-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.351,
      "max_context_k": 8,
      "layers": 46,
      "features": [],
      "mmlu_score": 75.2,
      "swe_bench_score": null,
      "notes": "46 layers, 16 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "gemma-2-27b-q8",
      "name": "Gemma 2 27B",
      "url": "https://whatmodelscanirun.com/model/gemma-2-27b",
      "quantization": "Q8_0",
      "params_b": 27.2,
      "weight_gb": 26.95,
      "gguf_repo": "bartowski/gemma-2-27b-it-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.351,
      "max_context_k": 8,
      "layers": 46,
      "features": [],
      "mmlu_score": 75.2,
      "swe_bench_score": null,
      "notes": "46 layers, 16 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "gemma-2-9b-q4",
      "name": "Gemma 2 9B",
      "url": "https://whatmodelscanirun.com/model/gemma-2-9b",
      "quantization": "Q4_K_M",
      "params_b": 9.24,
      "weight_gb": 5.37,
      "gguf_repo": "lmstudio-community/gemma-2-9b-it-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.321,
      "max_context_k": 8,
      "layers": 42,
      "features": [],
      "mmlu_score": 71.3,
      "swe_bench_score": null,
      "notes": "42 layers, 8 KV heads, head_dim 256. GGUF Q4_K_M file size."
    },
    {
      "id": "gemma-2-9b-q8",
      "name": "Gemma 2 9B",
      "url": "https://whatmodelscanirun.com/model/gemma-2-9b",
      "quantization": "Q8_0",
      "params_b": 9.24,
      "weight_gb": 9.15,
      "gguf_repo": "lmstudio-community/gemma-2-9b-it-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.321,
      "max_context_k": 8,
      "layers": 42,
      "features": [],
      "mmlu_score": 71.3,
      "swe_bench_score": null,
      "notes": "42 layers, 8 KV heads, head_dim 256. GGUF Q8_0 file size."
    },
    {
      "id": "gemma-3-12b-q4",
      "name": "Gemma 3 12B",
      "url": "https://whatmodelscanirun.com/model/gemma-3-12b",
      "quantization": "Q4_K_M",
      "params_b": 11.8,
      "weight_gb": 6.8,
      "gguf_repo": "unsloth/gemma-3-12b-it-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.061,
      "max_context_k": 128,
      "layers": 48,
      "features": [
        "vision"
      ],
      "mmlu_score": 74,
      "swe_bench_score": 12.8,
      "notes": "48 layers: 40 sliding-window (1024) and 8 full-attention (8 KV heads, head_dim 256), one full layer in every 6. Sliding layers hold a small fixed cache, so kv_per_1k_gb counts the full-attention layers only. GGUF Q4_K_M file size."
    },
    {
      "id": "gemma-3-12b-q8",
      "name": "Gemma 3 12B",
      "url": "https://whatmodelscanirun.com/model/gemma-3-12b",
      "quantization": "Q8_0",
      "params_b": 11.8,
      "weight_gb": 11.65,
      "gguf_repo": "unsloth/gemma-3-12b-it-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.061,
      "max_context_k": 128,
      "layers": 48,
      "features": [
        "vision"
      ],
      "mmlu_score": 74,
      "swe_bench_score": 12.8,
      "notes": "48 layers: 40 sliding-window (1024) and 8 full-attention (8 KV heads, head_dim 256), one full layer in every 6. Sliding layers hold a small fixed cache, so kv_per_1k_gb counts the full-attention layers only. GGUF Q8_0 file size."
    },
    {
      "id": "gemma-3-27b-q4",
      "name": "Gemma 3 27B",
      "url": "https://whatmodelscanirun.com/model/gemma-3-27b",
      "quantization": "Q4_K_M",
      "params_b": 27,
      "weight_gb": 15.41,
      "gguf_repo": "unsloth/gemma-3-27b-it-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.076,
      "max_context_k": 128,
      "layers": 62,
      "features": [
        "vision"
      ],
      "mmlu_score": 78,
      "swe_bench_score": 17.4,
      "notes": "62 layers: 52 sliding-window (1024) and 10 full-attention (16 KV heads, head_dim 128), one full layer in every 6. Sliding layers hold a small fixed cache, so kv_per_1k_gb counts the full-attention layers only. GGUF Q4_K_M file size."
    },
    {
      "id": "gemma-3-27b-q8",
      "name": "Gemma 3 27B",
      "url": "https://whatmodelscanirun.com/model/gemma-3-27b",
      "quantization": "Q8_0",
      "params_b": 27,
      "weight_gb": 26.74,
      "gguf_repo": "unsloth/gemma-3-27b-it-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.076,
      "max_context_k": 128,
      "layers": 62,
      "features": [
        "vision"
      ],
      "mmlu_score": 78,
      "swe_bench_score": 17.4,
      "notes": "62 layers: 52 sliding-window (1024) and 10 full-attention (16 KV heads, head_dim 128), one full layer in every 6. Sliding layers hold a small fixed cache, so kv_per_1k_gb counts the full-attention layers only. GGUF Q8_0 file size."
    },
    {
      "id": "gemma-3-4b-q4",
      "name": "Gemma 3 4B",
      "url": "https://whatmodelscanirun.com/model/gemma-3-4b",
      "quantization": "Q4_K_M",
      "params_b": 3.88,
      "weight_gb": 2.32,
      "gguf_repo": "unsloth/gemma-3-4b-it-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.019,
      "max_context_k": 128,
      "layers": 34,
      "features": [
        "vision"
      ],
      "mmlu_score": 63,
      "swe_bench_score": null,
      "notes": "34 layers: 29 sliding-window (1024) and 5 full-attention (4 KV heads, head_dim 256), one full layer in every 6. Sliding layers hold a small fixed cache, so kv_per_1k_gb counts the full-attention layers only. GGUF Q4_K_M file size."
    },
    {
      "id": "gemma-3-4b-q8",
      "name": "Gemma 3 4B",
      "url": "https://whatmodelscanirun.com/model/gemma-3-4b",
      "quantization": "Q8_0",
      "params_b": 3.88,
      "weight_gb": 3.85,
      "gguf_repo": "unsloth/gemma-3-4b-it-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.019,
      "max_context_k": 128,
      "layers": 34,
      "features": [
        "vision"
      ],
      "mmlu_score": 63,
      "swe_bench_score": null,
      "notes": "34 layers: 29 sliding-window (1024) and 5 full-attention (4 KV heads, head_dim 256), one full layer in every 6. Sliding layers hold a small fixed cache, so kv_per_1k_gb counts the full-attention layers only. GGUF Q8_0 file size."
    },
    {
      "id": "gemma-4-12b-q4",
      "name": "Gemma 4 12B",
      "url": "https://whatmodelscanirun.com/model/gemma-4-12b",
      "quantization": "Q4_K_M",
      "params_b": 11.96,
      "weight_gb": 6.63,
      "gguf_repo": "unsloth/gemma-4-12b-it-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.015,
      "max_context_k": 256,
      "layers": 48,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 77.2,
      "swe_bench_score": null,
      "notes": "Dense multimodal model. 48 layers: 40 sliding-window (1024) and 8 full-attention (1 KV head, global head_dim 512). 256K context. MMLU-Pro 77.2, ahead of Gemma 3 27B (67.6) at under half the VRAM. Sliding layers hold a small fixed cache, so kv_per_1k_gb counts the full-attention layers only. GGUF Q4_K_M file size."
    },
    {
      "id": "gemma-4-12b-q8",
      "name": "Gemma 4 12B",
      "url": "https://whatmodelscanirun.com/model/gemma-4-12b",
      "quantization": "Q8_0",
      "params_b": 11.96,
      "weight_gb": 11.8,
      "gguf_repo": "unsloth/gemma-4-12b-it-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.015,
      "max_context_k": 256,
      "layers": 48,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 77.2,
      "swe_bench_score": null,
      "notes": "Dense multimodal model. 48 layers: 40 sliding-window (1024) and 8 full-attention (1 KV head, global head_dim 512). 256K context. MMLU-Pro 77.2, ahead of Gemma 3 27B (67.6) at under half the VRAM. Sliding layers hold a small fixed cache, so kv_per_1k_gb counts the full-attention layers only. GGUF Q8_0 file size."
    },
    {
      "id": "gemma-4-26b-a4b-q4",
      "name": "Gemma 4 26B A4B",
      "url": "https://whatmodelscanirun.com/model/gemma-4-26b-a4b",
      "quantization": "Q4_K_M",
      "params_b": 26,
      "weight_gb": 15.87,
      "gguf_repo": "bartowski/google_gemma-4-26B-A4B-it-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 2.44,
      "kv_per_1k_gb": 0.019,
      "max_context_k": 256,
      "layers": 30,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 82.6,
      "swe_bench_score": null,
      "notes": "30 layers (5:1 local:global), 8 KV heads (local, sw=1024, head_dim 256), 2 KV heads (global, head_dim 512). Sparse MoE (128 experts, 8 active). KV per 1K from global attention only; local layers add ~0.20 GB fixed overhead. GGUF Q4_K_M file size."
    },
    {
      "id": "gemma-4-26b-a4b-q8",
      "name": "Gemma 4 26B A4B",
      "url": "https://whatmodelscanirun.com/model/gemma-4-26b-a4b",
      "quantization": "Q8_0",
      "params_b": 26,
      "weight_gb": 25.02,
      "gguf_repo": "unsloth/gemma-4-26B-A4B-it-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 3.85,
      "kv_per_1k_gb": 0.019,
      "max_context_k": 256,
      "layers": 30,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 82.6,
      "swe_bench_score": null,
      "notes": "30 layers (5:1 local:global), 8 KV heads (local, sw=1024, head_dim 256), 2 KV heads (global, head_dim 512). Sparse MoE (128 experts, 8 active). KV per 1K from global attention only; local layers add ~0.20 GB fixed overhead. GGUF Q8_0 file size."
    },
    {
      "id": "gemma-4-31b-q4",
      "name": "Gemma 4 31B",
      "url": "https://whatmodelscanirun.com/model/gemma-4-31b",
      "quantization": "Q4_K_M",
      "params_b": 31,
      "weight_gb": 17.07,
      "gguf_repo": "unsloth/gemma-4-31B-it-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.076,
      "max_context_k": 256,
      "layers": 60,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 85.2,
      "swe_bench_score": null,
      "notes": "60 layers (5:1 local:global), 16 KV heads (local, head_dim 256, sw=1024), 4 KV heads (global, head_dim 512). KV per 1K from global attention only; local layers add ~0.78 GB fixed overhead. GGUF Q4_K_M file size."
    },
    {
      "id": "gemma-4-31b-q8",
      "name": "Gemma 4 31B",
      "url": "https://whatmodelscanirun.com/model/gemma-4-31b",
      "quantization": "Q8_0",
      "params_b": 31,
      "weight_gb": 30.39,
      "gguf_repo": "unsloth/gemma-4-31B-it-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.076,
      "max_context_k": 256,
      "layers": 60,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 85.2,
      "swe_bench_score": null,
      "notes": "60 layers (5:1 local:global), 16 KV heads (local, head_dim 256, sw=1024), 4 KV heads (global, head_dim 512). KV per 1K from global attention only; local layers add ~0.78 GB fixed overhead. GGUF Q8_0 file size."
    },
    {
      "id": "gemma-4-e2b-q4",
      "name": "Gemma 4 E2B",
      "url": "https://whatmodelscanirun.com/model/gemma-4-e2b",
      "quantization": "Q4_K_M",
      "params_b": 2.3,
      "weight_gb": 3.22,
      "gguf_repo": "bartowski/google_gemma-4-E2B-it-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.006,
      "max_context_k": 128,
      "layers": 35,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 60,
      "swe_bench_score": null,
      "notes": "35 layers: 28 sliding-window (512) and 7 full-attention (1 KV head, global head_dim 512). The last 20 layers reuse earlier layers' KV, so 3 full-attention layers hold a growing cache. Multimodal MoE edge model. Sliding layers hold a small fixed cache, so kv_per_1k_gb counts the full-attention layers only. GGUF Q4_K_M file size."
    },
    {
      "id": "gemma-4-e2b-q8",
      "name": "Gemma 4 E2B",
      "url": "https://whatmodelscanirun.com/model/gemma-4-e2b",
      "quantization": "Q8_0",
      "params_b": 2.3,
      "weight_gb": 4.63,
      "gguf_repo": "bartowski/google_gemma-4-E2B-it-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.006,
      "max_context_k": 128,
      "layers": 35,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 60,
      "swe_bench_score": null,
      "notes": "35 layers: 28 sliding-window (512) and 7 full-attention (1 KV head, global head_dim 512). The last 20 layers reuse earlier layers' KV, so 3 full-attention layers hold a growing cache. Multimodal MoE edge model. Sliding layers hold a small fixed cache, so kv_per_1k_gb counts the full-attention layers only. GGUF Q8_0 file size."
    },
    {
      "id": "gemma-4-e4b-q4",
      "name": "Gemma 4 E4B",
      "url": "https://whatmodelscanirun.com/model/gemma-4-e4b",
      "quantization": "Q4_K_M",
      "params_b": 4.5,
      "weight_gb": 5.03,
      "gguf_repo": "bartowski/google_gemma-4-E4B-it-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.015,
      "max_context_k": 128,
      "layers": 42,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 69.4,
      "swe_bench_score": null,
      "notes": "42 layers: 35 sliding-window (512) and 7 full-attention (2 KV heads, global head_dim 512). The last 18 layers reuse earlier layers' KV, so 4 full-attention layers hold a growing cache. Multimodal MoE edge model. Sliding layers hold a small fixed cache, so kv_per_1k_gb counts the full-attention layers only. GGUF Q4_K_M file size."
    },
    {
      "id": "gemma-4-e4b-q8",
      "name": "Gemma 4 E4B",
      "url": "https://whatmodelscanirun.com/model/gemma-4-e4b",
      "quantization": "Q8_0",
      "params_b": 4.5,
      "weight_gb": 7.48,
      "gguf_repo": "bartowski/google_gemma-4-E4B-it-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.015,
      "max_context_k": 128,
      "layers": 42,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 69.4,
      "swe_bench_score": null,
      "notes": "42 layers: 35 sliding-window (512) and 7 full-attention (2 KV heads, global head_dim 512). The last 18 layers reuse earlier layers' KV, so 4 full-attention layers hold a growing cache. Multimodal MoE edge model. Sliding layers hold a small fixed cache, so kv_per_1k_gb counts the full-attention layers only. GGUF Q8_0 file size."
    },
    {
      "id": "glm-4.6-q4",
      "name": "GLM-4.6",
      "url": "https://whatmodelscanirun.com/model/glm-4-6",
      "quantization": "Q4_K_M",
      "params_b": 357,
      "weight_gb": 200.78,
      "gguf_repo": "unsloth/GLM-4.6-GGUF",
      "gguf_shards": 5,
      "active_weight_gb": 18,
      "kv_per_1k_gb": 0.067,
      "max_context_k": 200,
      "layers": 62,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 87,
      "swe_bench_score": 68.2,
      "notes": "MoE: 357B total. Expanded context to 200K from GLM-4.5. Uses MLA. KV estimate is for compressed MLA cache. GGUF Q4_K_M file size. Active params: 32B (GLM-4.5 model card; same architecture as GLM-4.5)."
    },
    {
      "id": "glm-4.6-q8",
      "name": "GLM-4.6",
      "url": "https://whatmodelscanirun.com/model/glm-4-6",
      "quantization": "Q8_0",
      "params_b": 357,
      "weight_gb": 353.27,
      "gguf_repo": "unsloth/GLM-4.6-GGUF",
      "gguf_shards": 8,
      "active_weight_gb": 31.67,
      "kv_per_1k_gb": 0.067,
      "max_context_k": 200,
      "layers": 62,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 87,
      "swe_bench_score": 68.2,
      "notes": "MoE: 357B total. Expanded context to 200K from GLM-4.5. Uses MLA. KV estimate is for compressed MLA cache. GGUF Q8_0 file size. Active params: 32B (GLM-4.5 model card; same architecture as GLM-4.5)."
    },
    {
      "id": "glm-4.7-flash-q4",
      "name": "GLM-4.7 Flash",
      "url": "https://whatmodelscanirun.com/model/glm-4-7-flash",
      "quantization": "Q4_K_M",
      "params_b": 31,
      "weight_gb": 17.05,
      "gguf_repo": "unsloth/GLM-4.7-Flash-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 1.65,
      "kv_per_1k_gb": 0.05,
      "max_context_k": 128,
      "layers": 47,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 82,
      "swe_bench_score": 59.2,
      "notes": "MoE: 30B total, 3B active (64 experts, 4 active). 47 layers, uses MLA (kv_lora_rank=512). KV estimate is for compressed MLA cache. GGUF Q4_K_M file size. Active params: 3B (GLM-4.7-Flash model card, 30B-A3B)."
    },
    {
      "id": "glm-4.7-flash-q8",
      "name": "GLM-4.7 Flash",
      "url": "https://whatmodelscanirun.com/model/glm-4-7-flash",
      "quantization": "Q8_0",
      "params_b": 31,
      "weight_gb": 29.66,
      "gguf_repo": "unsloth/GLM-4.7-Flash-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 2.87,
      "kv_per_1k_gb": 0.05,
      "max_context_k": 128,
      "layers": 47,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 82,
      "swe_bench_score": 59.2,
      "notes": "MoE: 30B total, 3B active (64 experts, 4 active). 47 layers, uses MLA (kv_lora_rank=512). KV estimate is for compressed MLA cache. GGUF Q8_0 file size. Active params: 3B (GLM-4.7-Flash model card, 30B-A3B)."
    },
    {
      "id": "glm-5-q4",
      "name": "GLM 5",
      "url": "https://whatmodelscanirun.com/model/glm-5",
      "quantization": "Q4_K_M",
      "params_b": 753.86,
      "weight_gb": 424.57,
      "gguf_repo": "unsloth/GLM-5-GGUF",
      "gguf_shards": 11,
      "active_weight_gb": 22.53,
      "kv_per_1k_gb": 0.084,
      "max_context_k": 198,
      "layers": 78,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 83,
      "swe_bench_score": 77.8,
      "notes": "Sparse MoE, 754B total (256 routed experts, 8 per token). 78 layers of Multi-head Latent Attention (kv_lora_rank 512 plus 64 RoPE dims). 198K context. SWE-bench Verified 77.8. GGUF Q4_K_M file size. Active params: 40B (GLM-5 model card)."
    },
    {
      "id": "glm-5-q8",
      "name": "GLM 5",
      "url": "https://whatmodelscanirun.com/model/glm-5",
      "quantization": "Q8_0",
      "params_b": 753.86,
      "weight_gb": 746.31,
      "gguf_repo": "unsloth/GLM-5-GGUF",
      "gguf_shards": 18,
      "active_weight_gb": 39.6,
      "kv_per_1k_gb": 0.084,
      "max_context_k": 198,
      "layers": 78,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 83,
      "swe_bench_score": 77.8,
      "notes": "Sparse MoE, 754B total (256 routed experts, 8 per token). 78 layers of Multi-head Latent Attention (kv_lora_rank 512 plus 64 RoPE dims). 198K context. SWE-bench Verified 77.8. GGUF Q8_0 file size. Active params: 40B (GLM-5 model card)."
    },
    {
      "id": "glm-5.1-q4",
      "name": "GLM 5.1",
      "url": "https://whatmodelscanirun.com/model/glm-5-1",
      "quantization": "Q4_K_M",
      "params_b": 754,
      "weight_gb": 428.11,
      "gguf_repo": "bartowski/zai-org_GLM-5.1-GGUF",
      "gguf_shards": 12,
      "active_weight_gb": 22.71,
      "kv_per_1k_gb": 0.084,
      "max_context_k": 198,
      "layers": 78,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 86.9,
      "swe_bench_score": null,
      "notes": "78 layers of Multi-head Latent Attention (kv_lora_rank 512 plus 64 RoPE dims), same as GLM 5. MoE 256 experts, 8 active. Active params: 40B (GLM-5 model card; same architecture as GLM-5)."
    },
    {
      "id": "glm-5.1-q8",
      "name": "GLM 5.1",
      "url": "https://whatmodelscanirun.com/model/glm-5-1",
      "quantization": "Q8_0",
      "params_b": 754,
      "weight_gb": 746.31,
      "gguf_repo": "unsloth/GLM-5.1-GGUF",
      "gguf_shards": 17,
      "active_weight_gb": 39.59,
      "kv_per_1k_gb": 0.084,
      "max_context_k": 198,
      "layers": 78,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 86.9,
      "swe_bench_score": null,
      "notes": "78 layers of Multi-head Latent Attention (kv_lora_rank 512 plus 64 RoPE dims), same as GLM 5. MoE 256 experts, 8 active. Active params: 40B (GLM-5 model card; same architecture as GLM-5)."
    },
    {
      "id": "glm-5.3-flash-q4",
      "name": "GLM 5.3 Flash",
      "url": "https://whatmodelscanirun.com/model/glm-5-3-flash",
      "quantization": "Q4_K_M",
      "params_b": 321.32,
      "weight_gb": 176.02,
      "gguf_repo": "DevQuasar/zai-org.GLM-5.3-Flash-GGUF",
      "gguf_shards": 15,
      "active_weight_gb": 9.86,
      "kv_per_1k_gb": 0.01,
      "max_context_k": 1024,
      "layers": 45,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 86.1,
      "swe_bench_score": null,
      "notes": "Sparse MoE (288 routed experts, 8 per token, plus 1 shared). 45 layers using Multi-head Latent Attention with kv_lora_rank 512: only 11 layers run DeepSeek-style sparse attention and cache a compressed latent, the other 34 are linear-attention, so KV cost per 1K tokens is unusually low for the size. Native 1M context. Terminal-Bench 2.1 84.3, DeepSWE v1.1 63.4. Z.ai did not publish MMLU-Pro or SWE-bench Verified. GGUF Q4_K_M file size. Active params: 18B (GLM-5.3-Flash model card)."
    },
    {
      "id": "glm-5.3-flash-q8",
      "name": "GLM 5.3 Flash",
      "url": "https://whatmodelscanirun.com/model/glm-5-3-flash",
      "quantization": "Q8_0",
      "params_b": 321.32,
      "weight_gb": 310.41,
      "gguf_repo": "DevQuasar/zai-org.GLM-5.3-Flash-GGUF",
      "gguf_shards": 26,
      "active_weight_gb": 17.39,
      "kv_per_1k_gb": 0.01,
      "max_context_k": 1024,
      "layers": 45,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 86.1,
      "swe_bench_score": null,
      "notes": "Sparse MoE (288 routed experts, 8 per token, plus 1 shared). 45 layers using Multi-head Latent Attention with kv_lora_rank 512: only 11 layers run DeepSeek-style sparse attention and cache a compressed latent, the other 34 are linear-attention, so KV cost per 1K tokens is unusually low for the size. Native 1M context. Terminal-Bench 2.1 84.3, DeepSWE v1.1 63.4. Z.ai did not publish MMLU-Pro or SWE-bench Verified. GGUF Q8_0 file size. Active params: 18B (GLM-5.3-Flash model card)."
    },
    {
      "id": "glm-5.3-q8",
      "name": "GLM 5.3",
      "url": "https://whatmodelscanirun.com/model/glm-5-3",
      "quantization": "Q8_0",
      "params_b": 753.33,
      "weight_gb": 746.32,
      "gguf_repo": "unsloth/GLM-5.3-GGUF",
      "gguf_shards": 17,
      "active_weight_gb": 39.63,
      "kv_per_1k_gb": 0.084,
      "max_context_k": 1024,
      "layers": 78,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 86.8,
      "swe_bench_score": null,
      "notes": "Z.ai flagship sparse MoE on the GLM-5 architecture: 753B total (256 routed experts, 8 per token, plus 1 shared, ~40B active). 78 layers of Multi-head Latent Attention with DeepSeek-style sparse attention (kv_lora_rank 512 plus 64 RoPE dims, one latent per token per layer). 1M context. DeepSWE v1.1 66.9, FrontierSWE 78.1. Z.ai did not publish MMLU-Pro or SWE-bench Verified. Only Q8_0 listed: no standard Q4_K_M from unsloth, bartowski or Z.ai (unsloth ships UD dynamic quants). GGUF Q8_0 file size."
    },
    {
      "id": "gpt-oss-20b-q4",
      "name": "GPT-OSS 20B",
      "url": "https://whatmodelscanirun.com/model/gpt-oss-20b",
      "quantization": "Q4_K_M",
      "params_b": 21,
      "weight_gb": 10.83,
      "gguf_repo": "unsloth/gpt-oss-20b-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 1.86,
      "kv_per_1k_gb": 0.023,
      "max_context_k": 128,
      "layers": 24,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 75,
      "swe_bench_score": 60.7,
      "notes": "MoE: 21B total, 3.6B active (32 experts, 4 per token). OpenAI open-weight model. 24 layers alternating 128-token sliding-window and full attention (8 KV heads, head_dim 64); only the 12 full-attention layers hold a growing cache. FFN weights stored as MXFP4. GGUF Q4_K_M file size. Active params: 3.6B (gpt-oss-20b model card)."
    },
    {
      "id": "gpt-oss-20b-q8",
      "name": "GPT-OSS 20B",
      "url": "https://whatmodelscanirun.com/model/gpt-oss-20b",
      "quantization": "Q8_0",
      "params_b": 21,
      "weight_gb": 11.28,
      "gguf_repo": "unsloth/gpt-oss-20b-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 1.93,
      "kv_per_1k_gb": 0.023,
      "max_context_k": 128,
      "layers": 24,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 75,
      "swe_bench_score": 60.7,
      "notes": "MoE: 21B total, 3.6B active (32 experts, 4 per token). OpenAI open-weight model. 24 layers alternating 128-token sliding-window and full attention (8 KV heads, head_dim 64); only the 12 full-attention layers hold a growing cache. FFN weights stored as MXFP4. GGUF Q8_0 file size. Active params: 3.6B (gpt-oss-20b model card)."
    },
    {
      "id": "k2-horizon-7b-q4",
      "name": "K2-Horizon 7B",
      "url": "https://whatmodelscanirun.com/model/k2-horizon-7b",
      "quantization": "Q4_K_M",
      "params_b": 9,
      "weight_gb": 5.21,
      "gguf_repo": "IFM/K2-Horizon-7B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.137,
      "max_context_k": 512,
      "layers": 36,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": 72.4,
      "notes": "IFM (MBZUAI) dense reasoning model, the medium member of the K2-Horizon family: 9.0B stored parameters around a 7B core. 36 full-attention layers (8 KV heads, head_dim 128), so KV cost grows quickly at its native 512K context. SWE-bench Verified 72.4, Terminal-Bench 2.1 39.1, BrowseComp 59.0, HMMT Feb 2026 73.3. MMLU-Pro not published. Apache 2.0. Official IFM GGUF. GGUF Q4_K_M file size."
    },
    {
      "id": "k2-horizon-7b-q8",
      "name": "K2-Horizon 7B",
      "url": "https://whatmodelscanirun.com/model/k2-horizon-7b",
      "quantization": "Q8_0",
      "params_b": 9,
      "weight_gb": 8.92,
      "gguf_repo": "IFM/K2-Horizon-7B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.137,
      "max_context_k": 512,
      "layers": 36,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": 72.4,
      "notes": "IFM (MBZUAI) dense reasoning model, the medium member of the K2-Horizon family: 9.0B stored parameters around a 7B core. 36 full-attention layers (8 KV heads, head_dim 128), so KV cost grows quickly at its native 512K context. SWE-bench Verified 72.4, Terminal-Bench 2.1 39.1, BrowseComp 59.0, HMMT Feb 2026 73.3. MMLU-Pro not published. Apache 2.0. Official IFM GGUF. GGUF Q8_0 file size."
    },
    {
      "id": "laguna-xs-2.1-q4",
      "name": "Laguna XS 2.1",
      "url": "https://whatmodelscanirun.com/model/laguna-xs-2-1",
      "quantization": "Q4_K_M",
      "params_b": 33.44,
      "weight_gb": 19.14,
      "gguf_repo": "bartowski/Laguna-XS-2.1-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 1.72,
      "kv_per_1k_gb": 0.038,
      "max_context_k": 256,
      "layers": 40,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": 70.9,
      "notes": "Agentic long-horizon coding model from poolside. Sparse MoE (256 experts, 8 routed per token, ~3B active). 40 layers, 10 full-attention (8 KV heads, head_dim 128), 30 sliding-window (512). 262K context. SWE-bench Verified 70.9. GGUF Q4_K_M file size."
    },
    {
      "id": "laguna-xs-2.1-q8",
      "name": "Laguna XS 2.1",
      "url": "https://whatmodelscanirun.com/model/laguna-xs-2-1",
      "quantization": "Q8_0",
      "params_b": 33.44,
      "weight_gb": 33.15,
      "gguf_repo": "bartowski/Laguna-XS-2.1-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 2.97,
      "kv_per_1k_gb": 0.038,
      "max_context_k": 256,
      "layers": 40,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": 70.9,
      "notes": "Agentic long-horizon coding model from poolside. Sparse MoE (256 experts, 8 routed per token, ~3B active). 40 layers, 10 full-attention (8 KV heads, head_dim 128), 30 sliding-window (512). 262K context. SWE-bench Verified 70.9. GGUF Q8_0 file size."
    },
    {
      "id": "lfm2.5-2.6b-q4",
      "name": "LFM2.5 2.6B",
      "url": "https://whatmodelscanirun.com/model/lfm2-5-2-6b",
      "quantization": "Q4_K_M",
      "params_b": 2.7,
      "weight_gb": 1.56,
      "gguf_repo": "LiquidAI/LFM2.5-2.6B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.015,
      "max_context_k": 128,
      "layers": 30,
      "features": [
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": null,
      "notes": "On-device agentic model. 30 layers, 8 full-attention (8 KV heads, head_dim 64) and 22 short-convolution layers, giving a very small KV footprint. 128K context. Strong instruction-following and tool-use scores for its size; MMLU not published for this release. GGUF Q4_K_M file size."
    },
    {
      "id": "lfm2.5-2.6b-q8",
      "name": "LFM2.5 2.6B",
      "url": "https://whatmodelscanirun.com/model/lfm2-5-2-6b",
      "quantization": "Q8_0",
      "params_b": 2.7,
      "weight_gb": 2.68,
      "gguf_repo": "LiquidAI/LFM2.5-2.6B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.015,
      "max_context_k": 128,
      "layers": 30,
      "features": [
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": null,
      "notes": "On-device agentic model. 30 layers, 8 full-attention (8 KV heads, head_dim 64) and 22 short-convolution layers, giving a very small KV footprint. 128K context. Strong instruction-following and tool-use scores for its size; MMLU not published for this release. GGUF Q8_0 file size."
    },
    {
      "id": "lfm2.5-24b-a2b-q4",
      "name": "LFM2.5 24B A2B",
      "url": "https://whatmodelscanirun.com/model/lfm2-5-24b-a2b",
      "quantization": "Q4_K_M",
      "params_b": 23.97,
      "weight_gb": 13.53,
      "gguf_repo": "LiquidAI/LFM2.5-24B-A2B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 1.38,
      "kv_per_1k_gb": 0.019,
      "max_context_k": 128,
      "layers": 40,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": 51,
      "notes": "Liquid AI's largest LFM2.5 model: sparse MoE (64 experts, 4 per token, 2.45B active) built for on-device agents. 40 layers: 30 short-convolution blocks plus 10 GQA attention layers (8 KV heads, head_dim 64), so only the attention layers hold a growing KV cache. 131K context. Always reasons before answering. SWE-bench Verified 51.0 (OpenCode harness). MMLU-Pro not published. Official LiquidAI GGUF. GGUF Q4_K_M file size."
    },
    {
      "id": "lfm2.5-24b-a2b-q8",
      "name": "LFM2.5 24B A2B",
      "url": "https://whatmodelscanirun.com/model/lfm2-5-24b-a2b",
      "quantization": "Q8_0",
      "params_b": 23.97,
      "weight_gb": 23.74,
      "gguf_repo": "LiquidAI/LFM2.5-24B-A2B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 2.43,
      "kv_per_1k_gb": 0.019,
      "max_context_k": 128,
      "layers": 40,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": 51,
      "notes": "Liquid AI's largest LFM2.5 model: sparse MoE (64 experts, 4 per token, 2.45B active) built for on-device agents. 40 layers: 30 short-convolution blocks plus 10 GQA attention layers (8 KV heads, head_dim 64), so only the attention layers hold a growing KV cache. 131K context. Always reasons before answering. SWE-bench Verified 51.0 (OpenCode harness). MMLU-Pro not published. Official LiquidAI GGUF. GGUF Q8_0 file size."
    },
    {
      "id": "lfm3-1.7b-q4",
      "name": "LFM3 1.7B",
      "url": "https://whatmodelscanirun.com/model/lfm3-1-7b",
      "quantization": "Q4_K_M",
      "params_b": 1.73,
      "weight_gb": 1.02,
      "gguf_repo": "LiquidAI/LFM3-1.7B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.011,
      "max_context_k": 256,
      "layers": 24,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": 43.4,
      "notes": "Liquid AI's on-device agent model. 24 layers: 13 gated short-convolution, 5 sliding-window (256 tokens) and 6 global attention (4 KV heads, head_dim 128), so only the 6 global layers grow the KV cache. 256K context. Reasons by default; built for tool calling, structured output and fine-tuning. SWE-bench Verified 43.4 (OpenCode harness). MMLU-Pro not published. Official LiquidAI GGUF. GGUF Q4_K_M file size."
    },
    {
      "id": "lfm3-1.7b-q8",
      "name": "LFM3 1.7B",
      "url": "https://whatmodelscanirun.com/model/lfm3-1-7b",
      "quantization": "Q8_0",
      "params_b": 1.73,
      "weight_gb": 1.72,
      "gguf_repo": "LiquidAI/LFM3-1.7B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.011,
      "max_context_k": 256,
      "layers": 24,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": 43.4,
      "notes": "Liquid AI's on-device agent model. 24 layers: 13 gated short-convolution, 5 sliding-window (256 tokens) and 6 global attention (4 KV heads, head_dim 128), so only the 6 global layers grow the KV cache. 256K context. Reasons by default; built for tool calling, structured output and fine-tuning. SWE-bench Verified 43.4 (OpenCode harness). MMLU-Pro not published. Official LiquidAI GGUF. GGUF Q8_0 file size."
    },
    {
      "id": "lfm3-vl-2b-q4",
      "name": "LFM3-VL 2B",
      "url": "https://whatmodelscanirun.com/model/lfm3-vl-2b",
      "quantization": "Q4_K_M",
      "params_b": 1.92,
      "weight_gb": 1.02,
      "gguf_repo": "LiquidAI/LFM3-VL-2B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.011,
      "max_context_k": 128,
      "layers": 24,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": null,
      "swe_bench_score": null,
      "notes": "Liquid AI's on-device vision-language model: the LFM3 1.7B language model plus a 164M vision encoder (1.92B total). Same 24-layer hybrid as LFM3 1.7B: 13 short-convolution, 5 sliding-window (256 tokens) and 6 global attention (4 KV heads, head_dim 128), so only the global layers grow the KV cache. 128K context. Thinks before answering; supports tool calls. MMLU-Pro and SWE-bench Verified not published. Weight excludes the separate mmproj file. Official LiquidAI GGUF. GGUF Q4_K_M file size."
    },
    {
      "id": "lfm3-vl-2b-q8",
      "name": "LFM3-VL 2B",
      "url": "https://whatmodelscanirun.com/model/lfm3-vl-2b",
      "quantization": "Q8_0",
      "params_b": 1.92,
      "weight_gb": 1.72,
      "gguf_repo": "LiquidAI/LFM3-VL-2B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.011,
      "max_context_k": 128,
      "layers": 24,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": null,
      "swe_bench_score": null,
      "notes": "Liquid AI's on-device vision-language model: the LFM3 1.7B language model plus a 164M vision encoder (1.92B total). Same 24-layer hybrid as LFM3 1.7B: 13 short-convolution, 5 sliding-window (256 tokens) and 6 global attention (4 KV heads, head_dim 128), so only the global layers grow the KV cache. 128K context. Thinks before answering; supports tool calls. MMLU-Pro and SWE-bench Verified not published. Weight excludes the separate mmproj file. Official LiquidAI GGUF. GGUF Q8_0 file size."
    },
    {
      "id": "ling-3.1-flash-q4",
      "name": "Ling 3.1 Flash",
      "url": "https://whatmodelscanirun.com/model/ling-3-1-flash",
      "quantization": "Q4_K_M",
      "params_b": 572.12,
      "weight_gb": 333.39,
      "gguf_repo": "bartowski/Ling-3.1-flash-1019-GGUF",
      "gguf_shards": 10,
      "active_weight_gb": 14.57,
      "kv_per_1k_gb": 0.01,
      "max_context_k": 256,
      "layers": 72,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": null,
      "notes": "inclusionAI's hybrid-linear sparse MoE, using the 1019 post-training refresh: ~560B total (512 routed experts, 8 per token, plus 1 shared), 25B active. 72 layers in a 7:1 pattern: 63 Kimi Delta Attention linear layers with fixed-size state and 9 gated MLA layers (kv_lora_rank 512 plus 64 RoPE dims), so only the 9 MLA layers grow the KV cache. 256K context as shipped in config.json and the GGUF; the card advertises 1M. Artificial Analysis Intelligence Index v5 30, AutomationBench-AA 68%, SciCode 57%. MMLU-Pro and SWE-bench Verified not published. MIT license. GGUF Q4_K_M file size."
    },
    {
      "id": "ling-3.1-flash-q8",
      "name": "Ling 3.1 Flash",
      "url": "https://whatmodelscanirun.com/model/ling-3-1-flash",
      "quantization": "Q8_0",
      "params_b": 572.12,
      "weight_gb": 566.6,
      "gguf_repo": "bartowski/Ling-3.1-flash-1019-GGUF",
      "gguf_shards": 16,
      "active_weight_gb": 24.76,
      "kv_per_1k_gb": 0.01,
      "max_context_k": 256,
      "layers": 72,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": null,
      "notes": "inclusionAI's hybrid-linear sparse MoE, using the 1019 post-training refresh: ~560B total (512 routed experts, 8 per token, plus 1 shared), 25B active. 72 layers in a 7:1 pattern: 63 Kimi Delta Attention linear layers with fixed-size state and 9 gated MLA layers (kv_lora_rank 512 plus 64 RoPE dims), so only the 9 MLA layers grow the KV cache. 256K context as shipped in config.json and the GGUF; the card advertises 1M. Artificial Analysis Intelligence Index v5 30, AutomationBench-AA 68%, SciCode 57%. MMLU-Pro and SWE-bench Verified not published. MIT license. GGUF Q8_0 file size."
    },
    {
      "id": "llama-3.1-8b-fp16",
      "name": "Llama 3.1 8B",
      "url": "https://whatmodelscanirun.com/model/llama-3-1-8b",
      "quantization": "fp16",
      "params_b": 8.03,
      "weight_gb": 14.96,
      "gguf_repo": null,
      "gguf_shards": null,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.122,
      "max_context_k": 128,
      "layers": 32,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 69.4,
      "swe_bench_score": 6.2,
      "notes": "32 layers, 8 KV heads, head_dim 128. Estimated from the GGUF f32 file size (bartowski, 29.92 GiB) / 2."
    },
    {
      "id": "llama-3.1-8b-q4",
      "name": "Llama 3.1 8B",
      "url": "https://whatmodelscanirun.com/model/llama-3-1-8b",
      "quantization": "Q4_K_M",
      "params_b": 8.03,
      "weight_gb": 4.58,
      "gguf_repo": "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.122,
      "max_context_k": 128,
      "layers": 32,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 69.4,
      "swe_bench_score": 6.2,
      "notes": "32 layers, 8 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "llama-3.1-8b-q8",
      "name": "Llama 3.1 8B",
      "url": "https://whatmodelscanirun.com/model/llama-3-1-8b",
      "quantization": "Q8_0",
      "params_b": 8.03,
      "weight_gb": 7.95,
      "gguf_repo": "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.122,
      "max_context_k": 128,
      "layers": 32,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 69.4,
      "swe_bench_score": 6.2,
      "notes": "32 layers, 8 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "llama-3.2-11b-vision-q4",
      "name": "Llama 3.2 Vision 11B",
      "url": "https://whatmodelscanirun.com/model/llama-3-2-vision-11b",
      "quantization": "Q4_K_M",
      "params_b": 10.6,
      "weight_gb": 5.96,
      "gguf_repo": null,
      "gguf_shards": null,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.122,
      "max_context_k": 128,
      "layers": 40,
      "features": [
        "tool_use",
        "vision"
      ],
      "mmlu_score": 73,
      "swe_bench_score": null,
      "notes": "Multimodal (vision adapter on Llama 3.1 8B). 40 layers (32 self-attn + 8 cross-attn), 8 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "llama-3.2-11b-vision-q8",
      "name": "Llama 3.2 Vision 11B",
      "url": "https://whatmodelscanirun.com/model/llama-3-2-vision-11b",
      "quantization": "Q8_0",
      "params_b": 10.6,
      "weight_gb": 10.4,
      "gguf_repo": null,
      "gguf_shards": null,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.122,
      "max_context_k": 128,
      "layers": 40,
      "features": [
        "tool_use",
        "vision"
      ],
      "mmlu_score": 73,
      "swe_bench_score": null,
      "notes": "Multimodal (vision adapter on Llama 3.1 8B). 40 layers (32 self-attn + 8 cross-attn), 8 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "llama-3.2-1b-q4",
      "name": "Llama 3.2 1B",
      "url": "https://whatmodelscanirun.com/model/llama-3-2-1b",
      "quantization": "Q4_K_M",
      "params_b": 1.24,
      "weight_gb": 0.75,
      "gguf_repo": "unsloth/Llama-3.2-1B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.031,
      "max_context_k": 128,
      "layers": 16,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 49.3,
      "swe_bench_score": null,
      "notes": "16 layers, 8 KV heads, head_dim 64. GGUF Q4_K_M file size."
    },
    {
      "id": "llama-3.2-1b-q8",
      "name": "Llama 3.2 1B",
      "url": "https://whatmodelscanirun.com/model/llama-3-2-1b",
      "quantization": "Q8_0",
      "params_b": 1.24,
      "weight_gb": 1.23,
      "gguf_repo": "unsloth/Llama-3.2-1B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.031,
      "max_context_k": 128,
      "layers": 16,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 49.3,
      "swe_bench_score": null,
      "notes": "16 layers, 8 KV heads, head_dim 64. GGUF Q8_0 file size."
    },
    {
      "id": "llama-3.2-3b-fp16",
      "name": "Llama 3.2 3B",
      "url": "https://whatmodelscanirun.com/model/llama-3-2-3b",
      "quantization": "fp16",
      "params_b": 3.21,
      "weight_gb": 5.99,
      "gguf_repo": "bartowski/Llama-3.2-3B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.107,
      "max_context_k": 128,
      "layers": 28,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 63.4,
      "swe_bench_score": null,
      "notes": "28 layers, 8 KV heads, head_dim 128. GGUF fp16 file size."
    },
    {
      "id": "llama-3.2-3b-q4",
      "name": "Llama 3.2 3B",
      "url": "https://whatmodelscanirun.com/model/llama-3-2-3b",
      "quantization": "Q4_K_M",
      "params_b": 3.21,
      "weight_gb": 1.88,
      "gguf_repo": "bartowski/Llama-3.2-3B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.107,
      "max_context_k": 128,
      "layers": 28,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 63.4,
      "swe_bench_score": null,
      "notes": "28 layers, 8 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "llama-3.2-3b-q8",
      "name": "Llama 3.2 3B",
      "url": "https://whatmodelscanirun.com/model/llama-3-2-3b",
      "quantization": "Q8_0",
      "params_b": 3.21,
      "weight_gb": 3.19,
      "gguf_repo": "bartowski/Llama-3.2-3B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.107,
      "max_context_k": 128,
      "layers": 28,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 63.4,
      "swe_bench_score": null,
      "notes": "28 layers, 8 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "llama-3.3-70b-q4",
      "name": "Llama 3.3 70B",
      "url": "https://whatmodelscanirun.com/model/llama-3-3-70b",
      "quantization": "Q4_K_M",
      "params_b": 70.6,
      "weight_gb": 39.6,
      "gguf_repo": "unsloth/Llama-3.3-70B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.305,
      "max_context_k": 128,
      "layers": 80,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 86,
      "swe_bench_score": 24.8,
      "notes": "80 layers, 8 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "llama-3.3-70b-q8",
      "name": "Llama 3.3 70B",
      "url": "https://whatmodelscanirun.com/model/llama-3-3-70b",
      "quantization": "Q8_0",
      "params_b": 70.6,
      "weight_gb": 69.83,
      "gguf_repo": "unsloth/Llama-3.3-70B-Instruct-GGUF",
      "gguf_shards": 2,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.305,
      "max_context_k": 128,
      "layers": 80,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 86,
      "swe_bench_score": 24.8,
      "notes": "80 layers, 8 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "llama-4-maverick-q4",
      "name": "Llama 4 Maverick",
      "url": "https://whatmodelscanirun.com/model/llama-4-maverick",
      "quantization": "Q4_K_M",
      "params_b": 400,
      "weight_gb": 226.09,
      "gguf_repo": "unsloth/Llama-4-Maverick-17B-128E-Instruct-GGUF",
      "gguf_shards": 5,
      "active_weight_gb": 9.61,
      "kv_per_1k_gb": 0.183,
      "max_context_k": 128,
      "layers": 48,
      "features": [
        "tool_use",
        "vision"
      ],
      "mmlu_score": 85.5,
      "swe_bench_score": null,
      "notes": "MoE: 400B total, 17B active (128 experts). 48 layers, 8 KV heads, head_dim 128. GGUF Q4_K_M file size. Active params: 17B (Meta Llama 4 announcement)."
    },
    {
      "id": "llama-4-maverick-q8",
      "name": "Llama 4 Maverick",
      "url": "https://whatmodelscanirun.com/model/llama-4-maverick",
      "quantization": "Q8_0",
      "params_b": 400,
      "weight_gb": 396.57,
      "gguf_repo": "unsloth/Llama-4-Maverick-17B-128E-Instruct-GGUF",
      "gguf_shards": 9,
      "active_weight_gb": 16.85,
      "kv_per_1k_gb": 0.183,
      "max_context_k": 128,
      "layers": 48,
      "features": [
        "tool_use",
        "vision"
      ],
      "mmlu_score": 85.5,
      "swe_bench_score": null,
      "notes": "MoE: 400B total, 17B active (128 experts). 48 layers, 8 KV heads, head_dim 128. GGUF Q8_0 file size. Active params: 17B (Meta Llama 4 announcement)."
    },
    {
      "id": "llama-4-scout-q4",
      "name": "Llama 4 Scout",
      "url": "https://whatmodelscanirun.com/model/llama-4-scout",
      "quantization": "Q4_K_M",
      "params_b": 109,
      "weight_gb": 60.87,
      "gguf_repo": "unsloth/Llama-4-Scout-17B-16E-Instruct-GGUF",
      "gguf_shards": 2,
      "active_weight_gb": 9.49,
      "kv_per_1k_gb": 0.183,
      "max_context_k": 128,
      "layers": 48,
      "features": [
        "tool_use",
        "vision"
      ],
      "mmlu_score": 79.6,
      "swe_bench_score": null,
      "notes": "MoE: 109B total, 17B active (16 experts). 48 layers, 8 KV heads, head_dim 128, hybrid attention (12 global + 36 local, chunk=8192). GGUF Q4_K_M file size. Active params: 17B (Meta Llama 4 announcement)."
    },
    {
      "id": "llama-4-scout-q8",
      "name": "Llama 4 Scout",
      "url": "https://whatmodelscanirun.com/model/llama-4-scout",
      "quantization": "Q8_0",
      "params_b": 109,
      "weight_gb": 106.67,
      "gguf_repo": "unsloth/Llama-4-Scout-17B-16E-Instruct-GGUF",
      "gguf_shards": 3,
      "active_weight_gb": 16.64,
      "kv_per_1k_gb": 0.183,
      "max_context_k": 128,
      "layers": 48,
      "features": [
        "tool_use",
        "vision"
      ],
      "mmlu_score": 79.6,
      "swe_bench_score": null,
      "notes": "MoE: 109B total, 17B active (16 experts). 48 layers, 8 KV heads, head_dim 128, hybrid attention (12 global + 36 local, chunk=8192). GGUF Q8_0 file size. Active params: 17B (Meta Llama 4 announcement)."
    },
    {
      "id": "magistral-small-24b-q4",
      "name": "Magistral Small 24B",
      "url": "https://whatmodelscanirun.com/model/magistral-small-24b",
      "quantization": "Q4_K_M",
      "params_b": 23.6,
      "weight_gb": 13.35,
      "gguf_repo": "mistralai/Magistral-Small-2509-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.153,
      "max_context_k": 128,
      "layers": 40,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 82,
      "swe_bench_score": null,
      "notes": "Reasoning model (Magistral 1.2). Based on Mistral Small 3.2. 40 layers, 8 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "magistral-small-24b-q8",
      "name": "Magistral Small 24B",
      "url": "https://whatmodelscanirun.com/model/magistral-small-24b",
      "quantization": "Q8_0",
      "params_b": 23.6,
      "weight_gb": 23.33,
      "gguf_repo": "mistralai/Magistral-Small-2509-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.153,
      "max_context_k": 128,
      "layers": 40,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 82,
      "swe_bench_score": null,
      "notes": "Reasoning model (Magistral 1.2). Based on Mistral Small 3.2. 40 layers, 8 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "mellum2.1-12b-a2.5b-thinking-q4",
      "name": "Mellum2.1 12B A2.5B Thinking",
      "url": "https://whatmodelscanirun.com/model/mellum2-1-12b-a2-5b-thinking",
      "quantization": "Q4_K_M",
      "params_b": 12.15,
      "weight_gb": 7.52,
      "gguf_repo": "JetBrains/Mellum2.1-12B-A2.5B-Thinking-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 1.55,
      "kv_per_1k_gb": 0.013,
      "max_context_k": 128,
      "layers": 28,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": 47,
      "notes": "JetBrains' coding-focused sparse MoE, 12B total with 2.5B active (64 experts, 8 per token), RL-trained for agentic software engineering. 28 layers: 7 full-attention (4 KV heads, head_dim 128) and 21 sliding-window (1,024 tokens), so only the full-attention layers grow the KV cache. 128K context. Always thinks before answering. SWE-bench Verified 47.0, SWE-bench Pro 28.0, LiveCodeBench v6 82.0, GPQA Diamond 64.6 (JetBrains self-reported, Pi harness). MMLU-Pro not published (MMLU-Redux 87.8). Weight excludes the MTP head files. Official JetBrains GGUF. Apache 2.0. GGUF Q4_K_M file size."
    },
    {
      "id": "mellum2.1-12b-a2.5b-thinking-q8",
      "name": "Mellum2.1 12B A2.5B Thinking",
      "url": "https://whatmodelscanirun.com/model/mellum2-1-12b-a2-5b-thinking",
      "quantization": "Q8_0",
      "params_b": 12.15,
      "weight_gb": 12.04,
      "gguf_repo": "JetBrains/Mellum2.1-12B-A2.5B-Thinking-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 2.48,
      "kv_per_1k_gb": 0.013,
      "max_context_k": 128,
      "layers": 28,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": 47,
      "notes": "JetBrains' coding-focused sparse MoE, 12B total with 2.5B active (64 experts, 8 per token), RL-trained for agentic software engineering. 28 layers: 7 full-attention (4 KV heads, head_dim 128) and 21 sliding-window (1,024 tokens), so only the full-attention layers grow the KV cache. 128K context. Always thinks before answering. SWE-bench Verified 47.0, SWE-bench Pro 28.0, LiveCodeBench v6 82.0, GPQA Diamond 64.6 (JetBrains self-reported, Pi harness). MMLU-Pro not published (MMLU-Redux 87.8). Weight excludes the MTP head files. Official JetBrains GGUF. Apache 2.0. GGUF Q8_0 file size."
    },
    {
      "id": "mimo-v2.5-pro-q4",
      "name": "MiMo V2.5 Pro",
      "url": "https://whatmodelscanirun.com/model/mimo-v2-5-pro",
      "quantization": "Q4_K_M",
      "params_b": 1020,
      "weight_gb": 586.37,
      "gguf_repo": "unsloth/MiMo-V2.5-Pro-GGUF",
      "gguf_shards": 15,
      "active_weight_gb": 24.14,
      "kv_per_1k_gb": 0.1,
      "max_context_k": 1024,
      "layers": 70,
      "features": [
        "reasoning"
      ],
      "mmlu_score": 89.4,
      "swe_bench_score": 78.9,
      "notes": "70 layers, 8 KV heads, head_dim 48; MoE 384 experts 8 active Active params: 42B (MiMo-V2.5-Pro model card)."
    },
    {
      "id": "mimo-v2.5-pro-q8",
      "name": "MiMo V2.5 Pro",
      "url": "https://whatmodelscanirun.com/model/mimo-v2-5-pro",
      "quantization": "Q8_0",
      "params_b": 1020,
      "weight_gb": 1012.92,
      "gguf_repo": "unsloth/MiMo-V2.5-Pro-GGUF",
      "gguf_shards": 24,
      "active_weight_gb": 41.71,
      "kv_per_1k_gb": 0.1,
      "max_context_k": 1024,
      "layers": 70,
      "features": [
        "reasoning"
      ],
      "mmlu_score": 89.4,
      "swe_bench_score": 78.9,
      "notes": "70 layers, 8 KV heads, head_dim 48; MoE 384 experts 8 active Active params: 42B (MiMo-V2.5-Pro model card)."
    },
    {
      "id": "mimo-v2.5-q4",
      "name": "MiMo V2.5",
      "url": "https://whatmodelscanirun.com/model/mimo-v2-5",
      "quantization": "Q4_K_M",
      "params_b": 310,
      "weight_gb": 175.8,
      "gguf_repo": "bartowski/MiMo-V2.5-GGUF",
      "gguf_shards": 5,
      "active_weight_gb": 8.51,
      "kv_per_1k_gb": 0.137,
      "max_context_k": 1024,
      "layers": 48,
      "features": [
        "reasoning"
      ],
      "mmlu_score": 82.9,
      "swe_bench_score": null,
      "notes": "48 layers, 4 KV heads, head_dim 192; MoE 256 experts 8 active Active params: 15B (MiMo-V2.5 model card)."
    },
    {
      "id": "mimo-v2.5-q8",
      "name": "MiMo V2.5",
      "url": "https://whatmodelscanirun.com/model/mimo-v2-5",
      "quantization": "Q8_0",
      "params_b": 310,
      "weight_gb": 306.67,
      "gguf_repo": "unsloth/MiMo-V2.5-GGUF",
      "gguf_shards": 8,
      "active_weight_gb": 14.84,
      "kv_per_1k_gb": 0.137,
      "max_context_k": 1024,
      "layers": 48,
      "features": [
        "reasoning"
      ],
      "mmlu_score": 82.9,
      "swe_bench_score": null,
      "notes": "48 layers, 4 KV heads, head_dim 192; MoE 256 experts 8 active Active params: 15B (MiMo-V2.5 model card)."
    },
    {
      "id": "mimo-v2.6-distill-qwen-9b-q4",
      "name": "MiMo V2.6 Distill Qwen 9B",
      "url": "https://whatmodelscanirun.com/model/mimo-v2-6-distill-qwen-9b",
      "quantization": "Q4_K_M",
      "params_b": 9.41,
      "weight_gb": 5.44,
      "gguf_repo": "bartowski/MiMo-V2.6-Distill-Qwen-9B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.031,
      "max_context_k": 256,
      "layers": 32,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": null,
      "swe_bench_score": 61.1,
      "notes": "Xiaomi MiMo's agentic SFT of Qwen3.5-9B on MiMo-V2.6-generated data, covering coding, general agent tasks and visual coding. Dense, 32 layers: 8 full-attention (4 KV heads, head_dim 256) and 24 linear-attention, so only the full-attention layers grow the KV cache. 262K context. SWE-bench Verified 61.1 (avg@3), SWE-bench Pro 44.6, Terminal-Bench 2.1 37.1. MMLU-Pro not published. Weight excludes the separate mmproj file. MIT license. GGUF Q4_K_M file size."
    },
    {
      "id": "mimo-v2.6-distill-qwen-9b-q8",
      "name": "MiMo V2.6 Distill Qwen 9B",
      "url": "https://whatmodelscanirun.com/model/mimo-v2-6-distill-qwen-9b",
      "quantization": "Q8_0",
      "params_b": 9.41,
      "weight_gb": 8.89,
      "gguf_repo": "bartowski/MiMo-V2.6-Distill-Qwen-9B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.031,
      "max_context_k": 256,
      "layers": 32,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": null,
      "swe_bench_score": 61.1,
      "notes": "Xiaomi MiMo's agentic SFT of Qwen3.5-9B on MiMo-V2.6-generated data, covering coding, general agent tasks and visual coding. Dense, 32 layers: 8 full-attention (4 KV heads, head_dim 256) and 24 linear-attention, so only the full-attention layers grow the KV cache. 262K context. SWE-bench Verified 61.1 (avg@3), SWE-bench Pro 44.6, Terminal-Bench 2.1 37.1. MMLU-Pro not published. Weight excludes the separate mmproj file. MIT license. GGUF Q8_0 file size."
    },
    {
      "id": "minicpm-v-4.7-35b-a3b-q4",
      "name": "MiniCPM-V 4.7 35B A3B",
      "url": "https://whatmodelscanirun.com/model/minicpm-v-4-7-35b-a3b",
      "quantization": "Q4_K_M",
      "params_b": 35.21,
      "weight_gb": 19.71,
      "gguf_repo": "openbmb/MiniCPM-V-4.7-35B-A3B-gguf",
      "gguf_shards": 1,
      "active_weight_gb": 1.68,
      "kv_per_1k_gb": 0.019,
      "max_context_k": 256,
      "layers": 40,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": null,
      "swe_bench_score": null,
      "notes": "OpenBMB vision-language model on a Qwen3.5/3.6-35B-A3B MoE backbone (256 experts, 8 routed, ~3B active) with an early-compression ViT supporting mixed 4x/16x visual token compression. 40 layers, 10 full-attention (2 KV heads, head_dim 256), 30 linear-attention. 262K context. Instruct and thinking modes, tool calling. MMLU-Pro and SWE-bench Verified not published. Weight excludes the separate mmproj file. Apache 2.0. Official openbmb GGUF. GGUF Q4_K_M file size."
    },
    {
      "id": "minicpm-v-4.7-35b-a3b-q8",
      "name": "MiniCPM-V 4.7 35B A3B",
      "url": "https://whatmodelscanirun.com/model/minicpm-v-4-7-35b-a3b",
      "quantization": "Q8_0",
      "params_b": 35.21,
      "weight_gb": 34.37,
      "gguf_repo": "openbmb/MiniCPM-V-4.7-35B-A3B-gguf",
      "gguf_shards": 1,
      "active_weight_gb": 2.93,
      "kv_per_1k_gb": 0.019,
      "max_context_k": 256,
      "layers": 40,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": null,
      "swe_bench_score": null,
      "notes": "OpenBMB vision-language model on a Qwen3.5/3.6-35B-A3B MoE backbone (256 experts, 8 routed, ~3B active) with an early-compression ViT supporting mixed 4x/16x visual token compression. 40 layers, 10 full-attention (2 KV heads, head_dim 256), 30 linear-attention. 262K context. Instruct and thinking modes, tool calling. MMLU-Pro and SWE-bench Verified not published. Weight excludes the separate mmproj file. Apache 2.0. Official openbmb GGUF. GGUF Q8_0 file size."
    },
    {
      "id": "minimax-m2.1-q4",
      "name": "MiniMax M2.1",
      "url": "https://whatmodelscanirun.com/model/minimax-m2-1",
      "quantization": "Q4_K_M",
      "params_b": 229,
      "weight_gb": 128.84,
      "gguf_repo": "unsloth/MiniMax-M2.1-GGUF",
      "gguf_shards": 3,
      "active_weight_gb": 5.63,
      "kv_per_1k_gb": 0.077,
      "max_context_k": 128,
      "layers": 80,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 88,
      "swe_bench_score": 74,
      "notes": "MoE: 229B total, ~10B active (128 experts, 2 per token). Hybrid lightning/softmax attention. KV estimate covers softmax attention layers only. GGUF Q4_K_M file size. Active params: 10B (MiniMax-M2 model card; same architecture as MiniMax-M2)."
    },
    {
      "id": "minimax-m2.1-q8",
      "name": "MiniMax M2.1",
      "url": "https://whatmodelscanirun.com/model/minimax-m2-1",
      "quantization": "Q8_0",
      "params_b": 229,
      "weight_gb": 226.44,
      "gguf_repo": "unsloth/MiniMax-M2.1-GGUF",
      "gguf_shards": 5,
      "active_weight_gb": 9.89,
      "kv_per_1k_gb": 0.077,
      "max_context_k": 128,
      "layers": 80,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 88,
      "swe_bench_score": 74,
      "notes": "MoE: 229B total, ~10B active (128 experts, 2 per token). Hybrid lightning/softmax attention. KV estimate covers softmax attention layers only. GGUF Q8_0 file size. Active params: 10B (MiniMax-M2 model card; same architecture as MiniMax-M2)."
    },
    {
      "id": "minimax-m2.5-q4",
      "name": "MiniMax M2.5",
      "url": "https://whatmodelscanirun.com/model/minimax-m2-5",
      "quantization": "Q4_K_M",
      "params_b": 228.7,
      "weight_gb": 128.84,
      "gguf_repo": "unsloth/MiniMax-M2.5-GGUF",
      "gguf_shards": 4,
      "active_weight_gb": 5.63,
      "kv_per_1k_gb": 0.237,
      "max_context_k": 192,
      "layers": 62,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 82,
      "swe_bench_score": 80.2,
      "notes": "Sparse MoE, 228.7B total, 8 experts routed per token. 62 layers of full attention (8 KV heads, head_dim 128), so the KV cache is large relative to the newer hybrid-attention models. 192K context. GGUF Q4_K_M file size. Active params: 10B (MiniMax-M2 model card; same architecture as MiniMax-M2)."
    },
    {
      "id": "minimax-m2.5-q8",
      "name": "MiniMax M2.5",
      "url": "https://whatmodelscanirun.com/model/minimax-m2-5",
      "quantization": "Q8_0",
      "params_b": 228.7,
      "weight_gb": 226.44,
      "gguf_repo": "unsloth/MiniMax-M2.5-GGUF",
      "gguf_shards": 6,
      "active_weight_gb": 9.9,
      "kv_per_1k_gb": 0.237,
      "max_context_k": 192,
      "layers": 62,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 82,
      "swe_bench_score": 80.2,
      "notes": "Sparse MoE, 228.7B total, 8 experts routed per token. 62 layers of full attention (8 KV heads, head_dim 128), so the KV cache is large relative to the newer hybrid-attention models. 192K context. GGUF Q8_0 file size. Active params: 10B (MiniMax-M2 model card; same architecture as MiniMax-M2)."
    },
    {
      "id": "ministral-8b-q4",
      "name": "Ministral 8B",
      "url": "https://whatmodelscanirun.com/model/ministral-8b",
      "quantization": "Q4_K_M",
      "params_b": 8.02,
      "weight_gb": 5.07,
      "gguf_repo": null,
      "gguf_shards": null,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.137,
      "max_context_k": 128,
      "layers": 36,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 65,
      "swe_bench_score": null,
      "notes": "Mistral edge model. 36 layers, 8 KV heads, head_dim 128. Interleaved sliding-window attention. GGUF Q4_K_M file size."
    },
    {
      "id": "ministral-8b-q8",
      "name": "Ministral 8B",
      "url": "https://whatmodelscanirun.com/model/ministral-8b",
      "quantization": "Q8_0",
      "params_b": 8.02,
      "weight_gb": 7.94,
      "gguf_repo": "bartowski/Ministral-8B-Instruct-2410-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.137,
      "max_context_k": 128,
      "layers": 36,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 65,
      "swe_bench_score": null,
      "notes": "Mistral edge model. 36 layers, 8 KV heads, head_dim 128. Interleaved sliding-window attention. GGUF Q8_0 file size."
    },
    {
      "id": "mistral-large-3-675b-q4",
      "name": "Mistral Large 3 675B",
      "url": "https://whatmodelscanirun.com/model/mistral-large-3-675b",
      "quantization": "Q4_K_M",
      "params_b": 673.42,
      "weight_gb": 379.04,
      "gguf_repo": "unsloth/Mistral-Large-3-675B-Instruct-2512-GGUF",
      "gguf_shards": 9,
      "active_weight_gb": 23.08,
      "kv_per_1k_gb": 0.065,
      "max_context_k": 288,
      "layers": 61,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 73.11,
      "swe_bench_score": null,
      "notes": "Granular MoE, 673B total with ~41B active per token, plus a 2.5B vision encoder. deepseek2 architecture: 61 layers of Multi-head Latent Attention (kv_lora_rank 512 plus 64 RoPE dims), so KV per 1K matches DeepSeek R1 despite the size. 288K context (294,912). Architecture read from the GGUF header; config.json is not published. MMLU-Pro 73.11, GPQA Diamond 67.17. Apache 2.0. GGUF Q4_K_M file size."
    },
    {
      "id": "mistral-large-3-675b-q8",
      "name": "Mistral Large 3 675B",
      "url": "https://whatmodelscanirun.com/model/mistral-large-3-675b",
      "quantization": "Q8_0",
      "params_b": 673.42,
      "weight_gb": 666.53,
      "gguf_repo": "unsloth/Mistral-Large-3-675B-Instruct-2512-GGUF",
      "gguf_shards": 15,
      "active_weight_gb": 40.58,
      "kv_per_1k_gb": 0.065,
      "max_context_k": 288,
      "layers": 61,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 73.11,
      "swe_bench_score": null,
      "notes": "Granular MoE, 673B total with ~41B active per token, plus a 2.5B vision encoder. deepseek2 architecture: 61 layers of Multi-head Latent Attention (kv_lora_rank 512 plus 64 RoPE dims), so KV per 1K matches DeepSeek R1 despite the size. 288K context (294,912). Architecture read from the GGUF header; config.json is not published. MMLU-Pro 73.11, GPQA Diamond 67.17. Apache 2.0. GGUF Q8_0 file size."
    },
    {
      "id": "mistral-large-4-1t-a52b-q4",
      "name": "Mistral Large 4 1T A52B",
      "url": "https://whatmodelscanirun.com/model/mistral-large-4-1t-a52b",
      "quantization": "Q4_K_M",
      "params_b": 1049.43,
      "weight_gb": 591.21,
      "gguf_repo": "ggml-org/Mistral-Large-4-1T-A52B-GGUF",
      "gguf_shards": 13,
      "active_weight_gb": 29.29,
      "kv_per_1k_gb": 0.068,
      "max_context_k": 1024,
      "layers": 72,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": null,
      "swe_bench_score": null,
      "notes": "Mistral's 1T-class granular MoE (128 routed experts, 4 per token, plus 1 shared; ~49B active, 52B including embeddings and output), natively multimodal with a separate vision encoder. 72 layers of Multi-head Latent Attention (kv_lora_rank 512 plus 64 RoPE dims, one latent per token per layer) with a DeepSeek-style sparse indexer; 9 layers reuse layer 35's KV cache, so only 63 allocate their own. 1M context. params_b derived from the BF16 GGUF size; config.json publishes no count. Artificial Analysis Coding Agent Index 48%, Terminal-Bench 4.0 41%, BrowseComp 84.5%. Mistral did not publish MMLU-Pro or SWE-bench Verified. Q4_K_M from ggml-org (unsloth ships only UD dynamic quants at 4-bit), Q8_0 from unsloth. Weight excludes the mmproj and DSpark draft files. GGUF Q4_K_M file size."
    },
    {
      "id": "mistral-large-4-1t-a52b-q8",
      "name": "Mistral Large 4 1T A52B",
      "url": "https://whatmodelscanirun.com/model/mistral-large-4-1t-a52b",
      "quantization": "Q8_0",
      "params_b": 1049.43,
      "weight_gb": 1038.56,
      "gguf_repo": "ggml-org/Mistral-Large-4-1T-A52B-GGUF",
      "gguf_shards": 24,
      "active_weight_gb": 51.46,
      "kv_per_1k_gb": 0.068,
      "max_context_k": 1024,
      "layers": 72,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": null,
      "swe_bench_score": null,
      "notes": "Mistral's 1T-class granular MoE (128 routed experts, 4 per token, plus 1 shared; ~49B active, 52B including embeddings and output), natively multimodal with a separate vision encoder. 72 layers of Multi-head Latent Attention (kv_lora_rank 512 plus 64 RoPE dims, one latent per token per layer) with a DeepSeek-style sparse indexer; 9 layers reuse layer 35's KV cache, so only 63 allocate their own. 1M context. params_b derived from the BF16 GGUF size; config.json publishes no count. Artificial Analysis Coding Agent Index 48%, Terminal-Bench 4.0 41%, BrowseComp 84.5%. Mistral did not publish MMLU-Pro or SWE-bench Verified. Q4_K_M from ggml-org (unsloth ships only UD dynamic quants at 4-bit), Q8_0 from unsloth. Weight excludes the mmproj and DSpark draft files. GGUF Q8_0 file size."
    },
    {
      "id": "mistral-nemo-12b-q4",
      "name": "Mistral Nemo 12B",
      "url": "https://whatmodelscanirun.com/model/mistral-nemo-12b",
      "quantization": "Q4_K_M",
      "params_b": 12.2,
      "weight_gb": 6.96,
      "gguf_repo": "unsloth/Mistral-Nemo-Instruct-2407-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.191,
      "max_context_k": 128,
      "layers": 40,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 68,
      "swe_bench_score": null,
      "notes": "Co-developed with NVIDIA. 40 layers, 8 KV heads, head_dim 160. GGUF Q4_K_M file size."
    },
    {
      "id": "mistral-nemo-12b-q8",
      "name": "Mistral Nemo 12B",
      "url": "https://whatmodelscanirun.com/model/mistral-nemo-12b",
      "quantization": "Q8_0",
      "params_b": 12.2,
      "weight_gb": 12.13,
      "gguf_repo": "unsloth/Mistral-Nemo-Instruct-2407-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.191,
      "max_context_k": 128,
      "layers": 40,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 68,
      "swe_bench_score": null,
      "notes": "Co-developed with NVIDIA. 40 layers, 8 KV heads, head_dim 160. GGUF Q8_0 file size."
    },
    {
      "id": "mistral-small-3.1-24b-q4",
      "name": "Mistral Small 3.1 24B",
      "url": "https://whatmodelscanirun.com/model/mistral-small-3-1-24b",
      "quantization": "Q4_K_M",
      "params_b": 23.6,
      "weight_gb": 13.35,
      "gguf_repo": "unsloth/Mistral-Small-3.1-24B-Instruct-2503-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.153,
      "max_context_k": 128,
      "layers": 40,
      "features": [
        "tool_use",
        "vision"
      ],
      "mmlu_score": 81,
      "swe_bench_score": 20.8,
      "notes": "40 layers, 8 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "mistral-small-3.1-24b-q8",
      "name": "Mistral Small 3.1 24B",
      "url": "https://whatmodelscanirun.com/model/mistral-small-3-1-24b",
      "quantization": "Q8_0",
      "params_b": 23.6,
      "weight_gb": 23.33,
      "gguf_repo": "unsloth/Mistral-Small-3.1-24B-Instruct-2503-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.153,
      "max_context_k": 128,
      "layers": 40,
      "features": [
        "tool_use",
        "vision"
      ],
      "mmlu_score": 81,
      "swe_bench_score": 20.8,
      "notes": "40 layers, 8 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "mistral-small-4-q4",
      "name": "Mistral Small 4",
      "url": "https://whatmodelscanirun.com/model/mistral-small-4",
      "quantization": "Q4_K_M",
      "params_b": 119,
      "weight_gb": 67.65,
      "gguf_repo": "bartowski/mistralai_Mistral-Small-4-119B-2603-GGUF",
      "gguf_shards": 2,
      "active_weight_gb": 3.7,
      "kv_per_1k_gb": 0.549,
      "max_context_k": 1024,
      "layers": 36,
      "features": [
        "tool_use",
        "vision"
      ],
      "mmlu_score": 88.5,
      "swe_bench_score": null,
      "notes": "36 layers, 32 KV heads, head_dim 128; MoE 128 experts 4 active; Pixtral vision Active params: 6.5B (Mistral-Small-4-119B-2603 model card)."
    },
    {
      "id": "mistral-small-4-q8",
      "name": "Mistral Small 4",
      "url": "https://whatmodelscanirun.com/model/mistral-small-4",
      "quantization": "Q8_0",
      "params_b": 119,
      "weight_gb": 117.79,
      "gguf_repo": "unsloth/Mistral-Small-4-119B-2603-GGUF",
      "gguf_shards": 4,
      "active_weight_gb": 6.43,
      "kv_per_1k_gb": 0.549,
      "max_context_k": 1024,
      "layers": 36,
      "features": [
        "tool_use",
        "vision"
      ],
      "mmlu_score": 88.5,
      "swe_bench_score": null,
      "notes": "36 layers, 32 KV heads, head_dim 128; MoE 128 experts 4 active; Pixtral vision Active params: 6.5B (Mistral-Small-4-119B-2603 model card)."
    },
    {
      "id": "muse-glimmer-30b-q4",
      "name": "Muse Glimmer 30B",
      "url": "https://whatmodelscanirun.com/model/muse-glimmer-30b",
      "quantization": "Q4_K_M",
      "params_b": 29.78,
      "weight_gb": 15.61,
      "gguf_repo": "meta-models/Muse-Glimmer-30B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.012,
      "max_context_k": 128,
      "layers": 52,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": null,
      "swe_bench_score": 76,
      "notes": "Meta's first open-weight release since Llama 4, distilled from Muse Spark. Dense multimodal: 27.8B language model plus a 1.8B frozen ViT vision encoder. 52 layers, 13 full-attention (2 KV heads, head_dim 128), 39 sliding-window (2048). 128K context. Weight excludes the separate mmproj vision projector file. SWE-bench Verified 76.0, SWE-bench Pro 51.2, AIME 2026 94.7, AA-GPQA Diamond 83.5. Apache 2.0. GGUF Q4_K_M file size."
    },
    {
      "id": "muse-glimmer-30b-q8",
      "name": "Muse Glimmer 30B",
      "url": "https://whatmodelscanirun.com/model/muse-glimmer-30b",
      "quantization": "Q8_0",
      "params_b": 29.78,
      "weight_gb": 27.58,
      "gguf_repo": "unsloth/Muse-Glimmer-30B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.012,
      "max_context_k": 128,
      "layers": 52,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": null,
      "swe_bench_score": 76,
      "notes": "Meta's first open-weight release since Llama 4, distilled from Muse Spark. Dense multimodal: 27.8B language model plus a 1.8B frozen ViT vision encoder. 52 layers, 13 full-attention (2 KV heads, head_dim 128), 39 sliding-window (2048). 128K context. Weight excludes the separate mmproj vision projector file. SWE-bench Verified 76.0, SWE-bench Pro 51.2, AIME 2026 94.7, AA-GPQA Diamond 83.5. Apache 2.0. GGUF Q8_0 file size."
    },
    {
      "id": "muse-spark-open-816b-a42b-q4",
      "name": "Muse Spark Open 816B A42B",
      "url": "https://whatmodelscanirun.com/model/muse-spark-open-816b-a42b",
      "quantization": "Q4_K_M",
      "params_b": 822.04,
      "weight_gb": 465.28,
      "gguf_repo": "DevQuasar/meta-models.Muse-Spark-Open-816B-A42B-GGUF",
      "gguf_shards": 39,
      "active_weight_gb": 23.77,
      "kv_per_1k_gb": 0.061,
      "max_context_k": 1024,
      "layers": 62,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": null,
      "swe_bench_score": null,
      "notes": "Meta's open-weight release of Muse Spark: natively multimodal MoE (256 experts, 8 per token, ~42B active) with a 1.8B Perception Encoder for image, video and audio input. 62 layers: 16 full-attention (16 KV heads, head_dim 64) and 46 sliding-window (2048) layers, so only the full-attention layers grow the KV cache. 1M context. Terminal-Bench 2.1 76.4, GPQA Diamond 89.0, DeepSWE 1.1 60. Meta did not publish MMLU-Pro or SWE-bench Verified. Q8_0 from ggml-org; Q4_K_M from DevQuasar (no standard Q4_K_M from unsloth or bartowski). Weight excludes the separate mmproj file. Muse Model License. GGUF Q4_K_M file size."
    },
    {
      "id": "muse-spark-open-816b-a42b-q8",
      "name": "Muse Spark Open 816B A42B",
      "url": "https://whatmodelscanirun.com/model/muse-spark-open-816b-a42b",
      "quantization": "Q8_0",
      "params_b": 822.04,
      "weight_gb": 811.79,
      "gguf_repo": "ggml-org/Muse-Spark-Open-816B-A42B-GGUF",
      "gguf_shards": 10,
      "active_weight_gb": 41.48,
      "kv_per_1k_gb": 0.061,
      "max_context_k": 1024,
      "layers": 62,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": null,
      "swe_bench_score": null,
      "notes": "Meta's open-weight release of Muse Spark: natively multimodal MoE (256 experts, 8 per token, ~42B active) with a 1.8B Perception Encoder for image, video and audio input. 62 layers: 16 full-attention (16 KV heads, head_dim 64) and 46 sliding-window (2048) layers, so only the full-attention layers grow the KV cache. 1M context. Terminal-Bench 2.1 76.4, GPQA Diamond 89.0, DeepSWE 1.1 60. Meta did not publish MMLU-Pro or SWE-bench Verified. Q8_0 from ggml-org; Q4_K_M from DevQuasar (no standard Q4_K_M from unsloth or bartowski). Weight excludes the separate mmproj file. Muse Model License. GGUF Q8_0 file size."
    },
    {
      "id": "nanbeige4.3-33b-q4",
      "name": "Nanbeige4.3 33B",
      "url": "https://whatmodelscanirun.com/model/nanbeige4-3-33b",
      "quantization": "Q4_K_M",
      "params_b": 33.12,
      "weight_gb": 19.36,
      "gguf_repo": "bartowski/Nanbeige_Nanbeige4.3-33B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.198,
      "max_context_k": 512,
      "layers": 52,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": null,
      "notes": "Dense 33B on the Llama architecture, flagship and distillation teacher of the Nanbeige4.3 series. 52 full-attention layers (8 KV heads, head_dim 128), so KV cost per 1K tokens is high next to hybrid models of similar size. 512K context. Thinking and tool-use modes. Terminal-Bench 2.1 80.9, DeepSWE 65.5, SWE-bench Pro 60.3, GPQA Diamond 89.5 (self-reported). MMLU-Pro and SWE-bench Verified not published. Apache 2.0. GGUF Q4_K_M file size."
    },
    {
      "id": "nanbeige4.3-33b-q8",
      "name": "Nanbeige4.3 33B",
      "url": "https://whatmodelscanirun.com/model/nanbeige4-3-33b",
      "quantization": "Q8_0",
      "params_b": 33.12,
      "weight_gb": 32.78,
      "gguf_repo": "bartowski/Nanbeige_Nanbeige4.3-33B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.198,
      "max_context_k": 512,
      "layers": 52,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": null,
      "notes": "Dense 33B on the Llama architecture, flagship and distillation teacher of the Nanbeige4.3 series. 52 full-attention layers (8 KV heads, head_dim 128), so KV cost per 1K tokens is high next to hybrid models of similar size. 512K context. Thinking and tool-use modes. Terminal-Bench 2.1 80.9, DeepSWE 65.5, SWE-bench Pro 60.3, GPQA Diamond 89.5 (self-reported). MMLU-Pro and SWE-bench Verified not published. Apache 2.0. GGUF Q8_0 file size."
    },
    {
      "id": "nemotron-3-nano-4b-q4",
      "name": "Nemotron 3 Nano 4B",
      "url": "https://whatmodelscanirun.com/model/nemotron-3-nano-4b",
      "quantization": "Q4_K_M",
      "params_b": 3.97,
      "weight_gb": 2.77,
      "gguf_repo": "bartowski/nvidia_Nemotron-3-Nano-4B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.015,
      "max_context_k": 256,
      "layers": 42,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": null,
      "notes": "Hybrid Mamba-Transformer (4 attention + 38 Mamba-2 layers). 8 KV heads, head_dim 128 on attention layers only. KV estimate covers attention layers; Mamba layers use fixed-size state. GGUF Q4_K_M file size."
    },
    {
      "id": "nemotron-3-nano-4b-q8",
      "name": "Nemotron 3 Nano 4B",
      "url": "https://whatmodelscanirun.com/model/nemotron-3-nano-4b",
      "quantization": "Q8_0",
      "params_b": 3.97,
      "weight_gb": 3.94,
      "gguf_repo": "bartowski/nvidia_Nemotron-3-Nano-4B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.015,
      "max_context_k": 256,
      "layers": 42,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": null,
      "notes": "Hybrid Mamba-Transformer (4 attention + 38 Mamba-2 layers). 8 KV heads, head_dim 128 on attention layers only. KV estimate covers attention layers; Mamba layers use fixed-size state. GGUF Q8_0 file size."
    },
    {
      "id": "nemotron-3.5-bolt-74b-a9b-q8",
      "name": "Nemotron 3.5 Bolt 74B A9B",
      "url": "https://whatmodelscanirun.com/model/nemotron-3-5-bolt-74b-a9b",
      "quantization": "Q8_0",
      "params_b": 74.89,
      "weight_gb": 73.44,
      "gguf_repo": "unsloth/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-GGUF",
      "gguf_shards": 3,
      "active_weight_gb": 8.83,
      "kv_per_1k_gb": 0.008,
      "max_context_k": 256,
      "layers": 88,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": null,
      "swe_bench_score": null,
      "notes": "Compressed (Puzzle/Star Elastic) derivative of Nemotron 3.5 Super VL: hybrid Mamba-Transformer MoE, 88 layers (40 Mamba-2, 40 MoE with top-9 routing, 8 attention), ~9B active per token, with the C-RADIO vision encoder. 2 KV heads, head_dim 128 on the 8 attention layers only. 256K context. NVIDIA's recommended local build is the official 24GB IQ2_XXS GGUF (22 GB); no standard Q4_K_M from a major quantizer yet. GPQA Diamond 88.8, Terminal-Bench 2.1 53.9. MMLU-Pro and SWE-bench Verified not published. Weight excludes the separate mmproj file. OpenMDW-1.1 license. GGUF Q8_0 file size."
    },
    {
      "id": "nemotron-3.5-bolt-e4b-q4",
      "name": "Nemotron 3.5 Bolt E4B",
      "url": "https://whatmodelscanirun.com/model/nemotron-3-5-bolt-e4b",
      "quantization": "Q4_K_M",
      "params_b": 8,
      "weight_gb": 5.76,
      "gguf_repo": "lmstudio-community/NVIDIA-Nemotron-3.5-Bolt-E4B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.015,
      "max_context_k": 64,
      "layers": 42,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": null,
      "notes": "NVIDIA post-train of Gemma 4 E4B for on-device assistants (text in, text out). 8.0B parameters stored including per-layer embeddings (~4B effective). 42 layers: 35 sliding-window (512) and 7 full-attention (2 KV heads, global head_dim 512); the last 18 layers reuse earlier KV caches, so only 4 full-attention layers grow the cache. 64K context. Thinking mode via chat template. Leads Gemma 4 E4B on MultiChallenge, IFBench and BFCL-v3. MMLU-Pro and SWE-bench Verified not published. GGUF Q4_K_M file size."
    },
    {
      "id": "nemotron-3.5-bolt-e4b-q8",
      "name": "Nemotron 3.5 Bolt E4B",
      "url": "https://whatmodelscanirun.com/model/nemotron-3-5-bolt-e4b",
      "quantization": "Q8_0",
      "params_b": 8,
      "weight_gb": 7.48,
      "gguf_repo": "lmstudio-community/NVIDIA-Nemotron-3.5-Bolt-E4B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.015,
      "max_context_k": 64,
      "layers": 42,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": null,
      "notes": "NVIDIA post-train of Gemma 4 E4B for on-device assistants (text in, text out). 8.0B parameters stored including per-layer embeddings (~4B effective). 42 layers: 35 sliding-window (512) and 7 full-attention (2 KV heads, global head_dim 512); the last 18 layers reuse earlier KV caches, so only 4 full-attention layers grow the cache. 64K context. Thinking mode via chat template. Leads Gemma 4 E4B on MultiChallenge, IFBench and BFCL-v3. MMLU-Pro and SWE-bench Verified not published. GGUF Q8_0 file size."
    },
    {
      "id": "nemotron-3.5-lightning-30b-a3b-q4",
      "name": "Nemotron 3.5 Lightning 30B A3B",
      "url": "https://whatmodelscanirun.com/model/nemotron-3-5-lightning-30b-a3b",
      "quantization": "Q4_K_M",
      "params_b": 31.58,
      "weight_gb": 23.73,
      "gguf_repo": "bartowski/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 2.25,
      "kv_per_1k_gb": 0.006,
      "max_context_k": 256,
      "layers": 52,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 81.94,
      "swe_bench_score": 51.56,
      "notes": "Hybrid Mamba-Transformer MoE (6 attention + 23 Mamba-2 + 23 MoE layers, 128 routed experts, 6 active, ~3B active per token). 2 KV heads, head_dim 128 on the attention layers only; Mamba layers hold fixed-size state. Config ships 262K context (up to 1M validated on large GPUs). MMLU-Pro 81.94, SWE-bench Verified 51.56. Weight excludes the separate MTP file. GGUF Q4_K_M file size."
    },
    {
      "id": "nemotron-3.5-lightning-30b-a3b-q8",
      "name": "Nemotron 3.5 Lightning 30B A3B",
      "url": "https://whatmodelscanirun.com/model/nemotron-3-5-lightning-30b-a3b",
      "quantization": "Q8_0",
      "params_b": 31.58,
      "weight_gb": 32.6,
      "gguf_repo": "unsloth/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 3.1,
      "kv_per_1k_gb": 0.006,
      "max_context_k": 256,
      "layers": 52,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 81.94,
      "swe_bench_score": 51.56,
      "notes": "Hybrid Mamba-Transformer MoE (6 attention + 23 Mamba-2 + 23 MoE layers, 128 routed experts, 6 active, ~3B active per token). 2 KV heads, head_dim 128 on the attention layers only; Mamba layers hold fixed-size state. Config ships 262K context (up to 1M validated on large GPUs). MMLU-Pro 81.94, SWE-bench Verified 51.56. Weight excludes the separate MTP file. GGUF Q8_0 file size."
    },
    {
      "id": "nemotron-3.5-super-vl-120b-a12b-q4",
      "name": "Nemotron 3.5 Super VL 120B A12B",
      "url": "https://whatmodelscanirun.com/model/nemotron-3-5-super-vl-120b-a12b",
      "quantization": "Q4_K_M",
      "params_b": 124.45,
      "weight_gb": 65.29,
      "gguf_repo": "lmstudio-community/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF",
      "gguf_shards": 2,
      "active_weight_gb": 6.3,
      "kv_per_1k_gb": 0.008,
      "max_context_k": 1024,
      "layers": 88,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": null,
      "swe_bench_score": 71.1,
      "notes": "Hybrid Mamba-Transformer MoE vision-language model: 88 layers (40 Mamba-2, 40 MoE with 512 experts / 22 active, 8 attention), ~12B active per token, plus a C-RADIO vision encoder. 2 KV heads, head_dim 128 on the 8 attention layers only, so the KV cache is very small. Up to 1M context. SWE-bench Verified 71.1, SWE-bench Pro 51.78, GPQA Diamond 89.5. Weight excludes the separate mmproj and MTP files. OpenMDW-1.1 license. GGUF Q4_K_M file size."
    },
    {
      "id": "nemotron-3.5-super-vl-120b-a12b-q8",
      "name": "Nemotron 3.5 Super VL 120B A12B",
      "url": "https://whatmodelscanirun.com/model/nemotron-3-5-super-vl-120b-a12b",
      "quantization": "Q8_0",
      "params_b": 124.45,
      "weight_gb": 119.65,
      "gguf_repo": "ggml-org/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 11.54,
      "kv_per_1k_gb": 0.008,
      "max_context_k": 1024,
      "layers": 88,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": null,
      "swe_bench_score": 71.1,
      "notes": "Hybrid Mamba-Transformer MoE vision-language model: 88 layers (40 Mamba-2, 40 MoE with 512 experts / 22 active, 8 attention), ~12B active per token, plus a C-RADIO vision encoder. 2 KV heads, head_dim 128 on the 8 attention layers only, so the KV cache is very small. Up to 1M context. SWE-bench Verified 71.1, SWE-bench Pro 51.78, GPQA Diamond 89.5. Weight excludes the separate mmproj and MTP files. OpenMDW-1.1 license. GGUF Q8_0 file size."
    },
    {
      "id": "nemotron-cascade-2-30b-a3b-q4",
      "name": "Nemotron Cascade 2 30B A3B",
      "url": "https://whatmodelscanirun.com/model/nemotron-cascade-2-30b-a3b",
      "quantization": "Q4_K_M",
      "params_b": 30,
      "weight_gb": 23.03,
      "gguf_repo": "bartowski/nvidia_Nemotron-Cascade-2-30B-A3B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 2.3,
      "kv_per_1k_gb": 0.006,
      "max_context_k": 256,
      "layers": 52,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 79.8,
      "swe_bench_score": 50.2,
      "notes": "Hybrid Mamba-Transformer MoE (6 attention + 23 Mamba-2 + 23 MoE layers, 128 experts, 6 active). 2 KV heads, head_dim 128 on attention layers only. KV estimate covers attention layers. GGUF Q4_K_M file size."
    },
    {
      "id": "nemotron-cascade-2-30b-a3b-q8",
      "name": "Nemotron Cascade 2 30B A3B",
      "url": "https://whatmodelscanirun.com/model/nemotron-cascade-2-30b-a3b",
      "quantization": "Q8_0",
      "params_b": 30,
      "weight_gb": 31.28,
      "gguf_repo": "bartowski/nvidia_Nemotron-Cascade-2-30B-A3B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 3.13,
      "kv_per_1k_gb": 0.006,
      "max_context_k": 256,
      "layers": 52,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 79.8,
      "swe_bench_score": 50.2,
      "notes": "Hybrid Mamba-Transformer MoE (6 attention + 23 Mamba-2 + 23 MoE layers, 128 experts, 6 active). 2 KV heads, head_dim 128 on attention layers only. KV estimate covers attention layers. GGUF Q8_0 file size."
    },
    {
      "id": "nemotron-nano-12b-q4",
      "name": "Nemotron Nano 12B",
      "url": "https://whatmodelscanirun.com/model/nemotron-nano-12b",
      "quantization": "Q4_K_M",
      "params_b": 12.6,
      "weight_gb": 6.98,
      "gguf_repo": "bartowski/nvidia_NVIDIA-Nemotron-Nano-12B-v2-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.023,
      "max_context_k": 128,
      "layers": 62,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 76,
      "swe_bench_score": null,
      "notes": "Hybrid Mamba-Transformer (6 attention + 56 Mamba-2 layers). 8 KV heads, head_dim 128 on attention layers only. KV estimate covers attention layers; Mamba layers use fixed-size state (~140 MB). GGUF Q4_K_M file size."
    },
    {
      "id": "nemotron-nano-12b-q8",
      "name": "Nemotron Nano 12B",
      "url": "https://whatmodelscanirun.com/model/nemotron-nano-12b",
      "quantization": "Q8_0",
      "params_b": 12.6,
      "weight_gb": 12.19,
      "gguf_repo": "bartowski/nvidia_NVIDIA-Nemotron-Nano-12B-v2-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.023,
      "max_context_k": 128,
      "layers": 62,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 76,
      "swe_bench_score": null,
      "notes": "Hybrid Mamba-Transformer (6 attention + 56 Mamba-2 layers). 8 KV heads, head_dim 128 on attention layers only. KV estimate covers attention layers; Mamba layers use fixed-size state (~140 MB). GGUF Q8_0 file size."
    },
    {
      "id": "ornith-1.5-122b-q4",
      "name": "Ornith 1.5 122B",
      "url": "https://whatmodelscanirun.com/model/ornith-1-5-122b",
      "quantization": "Q4_K_M",
      "params_b": 125.09,
      "weight_gb": 70.64,
      "gguf_repo": "ornith-ai/Ornith-1.5-122B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 5.65,
      "kv_per_1k_gb": 0.023,
      "max_context_k": 256,
      "layers": 48,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": 84.5,
      "notes": "Sparse MoE on the Qwen3.5-122B architecture (256 experts, 8 routed per token plus 1 shared, ~10B active). 48 layers, 12 full-attention (2 KV heads, head_dim 256), 36 linear-attention. 262K context. SWE-bench Verified 84.5 (OpenHands, anti-hack), SWE-bench Pro 76.9. MMLU-Pro not published. MIT license. Official ornith-ai GGUF. GGUF Q4_K_M file size."
    },
    {
      "id": "ornith-1.5-122b-q8",
      "name": "Ornith 1.5 122B",
      "url": "https://whatmodelscanirun.com/model/ornith-1-5-122b",
      "quantization": "Q8_0",
      "params_b": 125.09,
      "weight_gb": 123.45,
      "gguf_repo": "ornith-ai/Ornith-1.5-122B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 9.87,
      "kv_per_1k_gb": 0.023,
      "max_context_k": 256,
      "layers": 48,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": 84.5,
      "notes": "Sparse MoE on the Qwen3.5-122B architecture (256 experts, 8 routed per token plus 1 shared, ~10B active). 48 layers, 12 full-attention (2 KV heads, head_dim 256), 36 linear-attention. 262K context. SWE-bench Verified 84.5 (OpenHands, anti-hack), SWE-bench Pro 76.9. MMLU-Pro not published. MIT license. Official ornith-ai GGUF. GGUF Q8_0 file size."
    },
    {
      "id": "ornith-1.5-35b-a3b-q4",
      "name": "Ornith 1.5 35B A3B",
      "url": "https://whatmodelscanirun.com/model/ornith-1-5-35b-a3b",
      "quantization": "Q4_K_M",
      "params_b": 35.95,
      "weight_gb": 20.22,
      "gguf_repo": "ornith-ai/Ornith-1.5-35B-A3B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 1.69,
      "kv_per_1k_gb": 0.019,
      "max_context_k": 256,
      "layers": 40,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": 79,
      "notes": "Sparse MoE (256 experts, 8 routed per token, ~3B active). 40 layers, 10 full-attention (2 KV heads, head_dim 256), 30 linear-attention. 262K context. SWE-bench Verified 79.0, Terminal-Bench 2.1 67.8. MIT license. GGUF Q4_K_M file size."
    },
    {
      "id": "ornith-1.5-35b-a3b-q8",
      "name": "Ornith 1.5 35B A3B",
      "url": "https://whatmodelscanirun.com/model/ornith-1-5-35b-a3b",
      "quantization": "Q8_0",
      "params_b": 35.95,
      "weight_gb": 35.21,
      "gguf_repo": "ornith-ai/Ornith-1.5-35B-A3B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 2.94,
      "kv_per_1k_gb": 0.019,
      "max_context_k": 256,
      "layers": 40,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": 79,
      "notes": "Sparse MoE (256 experts, 8 routed per token, ~3B active). 40 layers, 10 full-attention (2 KV heads, head_dim 256), 30 linear-attention. 262K context. SWE-bench Verified 79.0, Terminal-Bench 2.1 67.8. MIT license. GGUF Q8_0 file size."
    },
    {
      "id": "ornith-1.5-397b-q4",
      "name": "Ornith 1.5 397B",
      "url": "https://whatmodelscanirun.com/model/ornith-1-5-397b",
      "quantization": "Q4_K_M",
      "params_b": 403.4,
      "weight_gb": 227.53,
      "gguf_repo": "ornith-ai/Ornith-1.5-397B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 9.02,
      "kv_per_1k_gb": 0.029,
      "max_context_k": 256,
      "layers": 60,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": 86,
      "notes": "Sparse MoE (512 experts, 10 routed per token, ~16B active). 60 layers, 15 full-attention (2 KV heads, head_dim 256), 45 linear-attention. 262K context. SWE-bench Verified 86.0, Terminal-Bench 2.1 86.1, WideSearch 80.8 - competitive with frontier closed models on agentic coding. MIT license. GGUF Q4_K_M file size."
    },
    {
      "id": "ornith-1.5-397b-q8",
      "name": "Ornith 1.5 397B",
      "url": "https://whatmodelscanirun.com/model/ornith-1-5-397b",
      "quantization": "Q8_0",
      "params_b": 403.4,
      "weight_gb": 399.09,
      "gguf_repo": "ornith-ai/Ornith-1.5-397B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 15.83,
      "kv_per_1k_gb": 0.029,
      "max_context_k": 256,
      "layers": 60,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": 86,
      "notes": "Sparse MoE (512 experts, 10 routed per token, ~16B active). 60 layers, 15 full-attention (2 KV heads, head_dim 256), 45 linear-attention. 262K context. SWE-bench Verified 86.0, Terminal-Bench 2.1 86.1, WideSearch 80.8 - competitive with frontier closed models on agentic coding. MIT license. GGUF Q8_0 file size."
    },
    {
      "id": "ornith-1.5-9b-q4",
      "name": "Ornith 1.5 9B",
      "url": "https://whatmodelscanirun.com/model/ornith-1-5-9b",
      "quantization": "Q4_K_M",
      "params_b": 9.65,
      "weight_gb": 5.38,
      "gguf_repo": "ornith-ai/Ornith-1.5-9B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.031,
      "max_context_k": 256,
      "layers": 32,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": 70.6,
      "notes": "Dense model trained with a self-improvement loop. 32 layers, 8 full-attention (4 KV heads, head_dim 256), 24 linear-attention. 262K context. SWE-bench Verified 70.6, Terminal-Bench 2.1 46.2. MIT license. GGUF Q4_K_M file size."
    },
    {
      "id": "ornith-1.5-9b-q8",
      "name": "Ornith 1.5 9B",
      "url": "https://whatmodelscanirun.com/model/ornith-1-5-9b",
      "quantization": "Q8_0",
      "params_b": 9.65,
      "weight_gb": 9.11,
      "gguf_repo": "ornith-ai/Ornith-1.5-9B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.031,
      "max_context_k": 256,
      "layers": 32,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": null,
      "swe_bench_score": 70.6,
      "notes": "Dense model trained with a self-improvement loop. 32 layers, 8 full-attention (4 KV heads, head_dim 256), 24 linear-attention. 262K context. SWE-bench Verified 70.6, Terminal-Bench 2.1 46.2. MIT license. GGUF Q8_0 file size."
    },
    {
      "id": "phi-3-mini-3.8b-q4",
      "name": "Phi-3 Mini 3.8B",
      "url": "https://whatmodelscanirun.com/model/phi-3-mini-3-8b",
      "quantization": "Q4_K_M",
      "params_b": 3.82,
      "weight_gb": 2.23,
      "gguf_repo": "bartowski/Phi-3.1-mini-128k-instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.366,
      "max_context_k": 128,
      "layers": 32,
      "features": [],
      "mmlu_score": 70.9,
      "swe_bench_score": null,
      "notes": "32 layers, 32 KV heads (full MHA), head_dim 96. GGUF Q4_K_M file size."
    },
    {
      "id": "phi-3-mini-3.8b-q8",
      "name": "Phi-3 Mini 3.8B",
      "url": "https://whatmodelscanirun.com/model/phi-3-mini-3-8b",
      "quantization": "Q8_0",
      "params_b": 3.82,
      "weight_gb": 3.78,
      "gguf_repo": "bartowski/Phi-3.1-mini-128k-instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.366,
      "max_context_k": 128,
      "layers": 32,
      "features": [],
      "mmlu_score": 70.9,
      "swe_bench_score": null,
      "notes": "32 layers, 32 KV heads (full MHA), head_dim 96. GGUF Q8_0 file size."
    },
    {
      "id": "phi-4-14b-q4",
      "name": "Phi-4 14B",
      "url": "https://whatmodelscanirun.com/model/phi-4-14b",
      "quantization": "Q4_K_M",
      "params_b": 14.7,
      "weight_gb": 8.43,
      "gguf_repo": "bartowski/phi-4-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.191,
      "max_context_k": 16,
      "layers": 40,
      "features": [],
      "mmlu_score": 84.8,
      "swe_bench_score": 18.6,
      "notes": "40 layers, 10 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "phi-4-14b-q8",
      "name": "Phi-4 14B",
      "url": "https://whatmodelscanirun.com/model/phi-4-14b",
      "quantization": "Q8_0",
      "params_b": 14.7,
      "weight_gb": 14.51,
      "gguf_repo": "unsloth/phi-4-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.191,
      "max_context_k": 16,
      "layers": 40,
      "features": [],
      "mmlu_score": 84.8,
      "swe_bench_score": 18.6,
      "notes": "40 layers, 10 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "phi-4-mini-3.8b-q4",
      "name": "Phi-4 Mini 3.8B",
      "url": "https://whatmodelscanirun.com/model/phi-4-mini-3-8b",
      "quantization": "Q4_K_M",
      "params_b": 3.84,
      "weight_gb": 2.32,
      "gguf_repo": "unsloth/Phi-4-mini-instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.122,
      "max_context_k": 128,
      "layers": 32,
      "features": [],
      "mmlu_score": 70,
      "swe_bench_score": null,
      "notes": "32 layers, 8 KV heads, head_dim 128. 200K vocabulary. GGUF Q4_K_M file size."
    },
    {
      "id": "phi-4-mini-3.8b-q8",
      "name": "Phi-4 Mini 3.8B",
      "url": "https://whatmodelscanirun.com/model/phi-4-mini-3-8b",
      "quantization": "Q8_0",
      "params_b": 3.84,
      "weight_gb": 3.8,
      "gguf_repo": "unsloth/Phi-4-mini-instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.122,
      "max_context_k": 128,
      "layers": 32,
      "features": [],
      "mmlu_score": 70,
      "swe_bench_score": null,
      "notes": "32 layers, 8 KV heads, head_dim 128. 200K vocabulary. GGUF Q8_0 file size."
    },
    {
      "id": "qwen-2.5-14b-q4",
      "name": "Qwen 2.5 14B",
      "url": "https://whatmodelscanirun.com/model/qwen-2-5-14b",
      "quantization": "Q4_K_M",
      "params_b": 14.7,
      "weight_gb": 8.37,
      "gguf_repo": "bartowski/Qwen2.5-14B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.183,
      "max_context_k": 128,
      "layers": 48,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 79.9,
      "swe_bench_score": 17.8,
      "notes": "48 layers, 8 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "qwen-2.5-14b-q8",
      "name": "Qwen 2.5 14B",
      "url": "https://whatmodelscanirun.com/model/qwen-2-5-14b",
      "quantization": "Q8_0",
      "params_b": 14.7,
      "weight_gb": 14.62,
      "gguf_repo": "bartowski/Qwen2.5-14B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.183,
      "max_context_k": 128,
      "layers": 48,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 79.9,
      "swe_bench_score": 17.8,
      "notes": "48 layers, 8 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "qwen-2.5-32b-q4",
      "name": "Qwen 2.5 32B",
      "url": "https://whatmodelscanirun.com/model/qwen-2-5-32b",
      "quantization": "Q4_K_M",
      "params_b": 32.5,
      "weight_gb": 18.49,
      "gguf_repo": "bartowski/Qwen2.5-32B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.244,
      "max_context_k": 128,
      "layers": 64,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 83.3,
      "swe_bench_score": 24.2,
      "notes": "64 layers, 8 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "qwen-2.5-32b-q8",
      "name": "Qwen 2.5 32B",
      "url": "https://whatmodelscanirun.com/model/qwen-2-5-32b",
      "quantization": "Q8_0",
      "params_b": 32.5,
      "weight_gb": 32.43,
      "gguf_repo": "bartowski/Qwen2.5-32B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.244,
      "max_context_k": 128,
      "layers": 64,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 83.3,
      "swe_bench_score": 24.2,
      "notes": "64 layers, 8 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "qwen-2.5-3b-q4",
      "name": "Qwen 2.5 3B",
      "url": "https://whatmodelscanirun.com/model/qwen-2-5-3b",
      "quantization": "Q4_K_M",
      "params_b": 3.09,
      "weight_gb": 1.8,
      "gguf_repo": "bartowski/Qwen2.5-3B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.034,
      "max_context_k": 32,
      "layers": 36,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 65.6,
      "swe_bench_score": null,
      "notes": "36 layers, 2 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "qwen-2.5-3b-q8",
      "name": "Qwen 2.5 3B",
      "url": "https://whatmodelscanirun.com/model/qwen-2-5-3b",
      "quantization": "Q8_0",
      "params_b": 3.09,
      "weight_gb": 3.06,
      "gguf_repo": "bartowski/Qwen2.5-3B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.034,
      "max_context_k": 32,
      "layers": 36,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 65.6,
      "swe_bench_score": null,
      "notes": "36 layers, 2 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "qwen-2.5-72b-q4",
      "name": "Qwen 2.5 72B",
      "url": "https://whatmodelscanirun.com/model/qwen-2-5-72b",
      "quantization": "Q4_K_M",
      "params_b": 72.7,
      "weight_gb": 44.16,
      "gguf_repo": "bartowski/Qwen2.5-72B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.305,
      "max_context_k": 128,
      "layers": 80,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 86.1,
      "swe_bench_score": 28.4,
      "notes": "80 layers, 8 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "qwen-2.5-72b-q8",
      "name": "Qwen 2.5 72B",
      "url": "https://whatmodelscanirun.com/model/qwen-2-5-72b",
      "quantization": "Q8_0",
      "params_b": 72.7,
      "weight_gb": 71.96,
      "gguf_repo": "bartowski/Qwen2.5-72B-Instruct-GGUF",
      "gguf_shards": 2,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.305,
      "max_context_k": 128,
      "layers": 80,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 86.1,
      "swe_bench_score": 28.4,
      "notes": "80 layers, 8 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "qwen-2.5-7b-q4",
      "name": "Qwen 2.5 7B",
      "url": "https://whatmodelscanirun.com/model/qwen-2-5-7b",
      "quantization": "Q4_K_M",
      "params_b": 7.62,
      "weight_gb": 4.36,
      "gguf_repo": "bartowski/Qwen2.5-7B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.053,
      "max_context_k": 128,
      "layers": 28,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 74.2,
      "swe_bench_score": 11.6,
      "notes": "28 layers, 4 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "qwen-2.5-7b-q8",
      "name": "Qwen 2.5 7B",
      "url": "https://whatmodelscanirun.com/model/qwen-2-5-7b",
      "quantization": "Q8_0",
      "params_b": 7.62,
      "weight_gb": 7.54,
      "gguf_repo": "bartowski/Qwen2.5-7B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.053,
      "max_context_k": 128,
      "layers": 28,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 74.2,
      "swe_bench_score": 11.6,
      "notes": "28 layers, 4 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "qwen-2.5-coder-14b-q4",
      "name": "Qwen 2.5 Coder 14B",
      "url": "https://whatmodelscanirun.com/model/qwen-2-5-coder-14b",
      "quantization": "Q4_K_M",
      "params_b": 14.7,
      "weight_gb": 8.37,
      "gguf_repo": "unsloth/Qwen2.5-Coder-14B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.183,
      "max_context_k": 128,
      "layers": 48,
      "features": [],
      "mmlu_score": 74,
      "swe_bench_score": 22.6,
      "notes": "Code model. Based on Qwen 2.5 14B. 48 layers, 8 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "qwen-2.5-coder-14b-q8",
      "name": "Qwen 2.5 Coder 14B",
      "url": "https://whatmodelscanirun.com/model/qwen-2-5-coder-14b",
      "quantization": "Q8_0",
      "params_b": 14.7,
      "weight_gb": 14.62,
      "gguf_repo": "unsloth/Qwen2.5-Coder-14B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.183,
      "max_context_k": 128,
      "layers": 48,
      "features": [],
      "mmlu_score": 74,
      "swe_bench_score": 22.6,
      "notes": "Code model. Based on Qwen 2.5 14B. 48 layers, 8 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "qwen-2.5-coder-32b-q4",
      "name": "Qwen 2.5 Coder 32B",
      "url": "https://whatmodelscanirun.com/model/qwen-2-5-coder-32b",
      "quantization": "Q4_K_M",
      "params_b": 32.5,
      "weight_gb": 18.49,
      "gguf_repo": "unsloth/Qwen2.5-Coder-32B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.244,
      "max_context_k": 128,
      "layers": 64,
      "features": [],
      "mmlu_score": 79,
      "swe_bench_score": 33.4,
      "notes": "Code model. Based on Qwen 2.5 32B. 64 layers, 8 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "qwen-2.5-coder-32b-q8",
      "name": "Qwen 2.5 Coder 32B",
      "url": "https://whatmodelscanirun.com/model/qwen-2-5-coder-32b",
      "quantization": "Q8_0",
      "params_b": 32.5,
      "weight_gb": 32.43,
      "gguf_repo": "unsloth/Qwen2.5-Coder-32B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.244,
      "max_context_k": 128,
      "layers": 64,
      "features": [],
      "mmlu_score": 79,
      "swe_bench_score": 33.4,
      "notes": "Code model. Based on Qwen 2.5 32B. 64 layers, 8 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "qwen-2.5-coder-7b-q4",
      "name": "Qwen 2.5 Coder 7B",
      "url": "https://whatmodelscanirun.com/model/qwen-2-5-coder-7b",
      "quantization": "Q4_K_M",
      "params_b": 7.62,
      "weight_gb": 4.36,
      "gguf_repo": "unsloth/Qwen2.5-Coder-7B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.053,
      "max_context_k": 128,
      "layers": 28,
      "features": [],
      "mmlu_score": 68,
      "swe_bench_score": 14.8,
      "notes": "Code model. Based on Qwen 2.5 7B. 28 layers, 4 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "qwen-2.5-coder-7b-q8",
      "name": "Qwen 2.5 Coder 7B",
      "url": "https://whatmodelscanirun.com/model/qwen-2-5-coder-7b",
      "quantization": "Q8_0",
      "params_b": 7.62,
      "weight_gb": 7.54,
      "gguf_repo": "unsloth/Qwen2.5-Coder-7B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.053,
      "max_context_k": 128,
      "layers": 28,
      "features": [],
      "mmlu_score": 68,
      "swe_bench_score": 14.8,
      "notes": "Code model. Based on Qwen 2.5 7B. 28 layers, 4 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "qwen-2.5-vl-3b-q4",
      "name": "Qwen 2.5 VL 3B",
      "url": "https://whatmodelscanirun.com/model/qwen-2-5-vl-3b",
      "quantization": "Q4_K_M",
      "params_b": 3.09,
      "weight_gb": 1.8,
      "gguf_repo": "unsloth/Qwen2.5-VL-3B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.034,
      "max_context_k": 128,
      "layers": 36,
      "features": [
        "vision"
      ],
      "mmlu_score": 65.6,
      "swe_bench_score": null,
      "notes": "Vision-language model. Same text backbone as Qwen 2.5 3B. 36 layers, 2 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "qwen-2.5-vl-3b-q8",
      "name": "Qwen 2.5 VL 3B",
      "url": "https://whatmodelscanirun.com/model/qwen-2-5-vl-3b",
      "quantization": "Q8_0",
      "params_b": 3.09,
      "weight_gb": 3.06,
      "gguf_repo": "unsloth/Qwen2.5-VL-3B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.034,
      "max_context_k": 128,
      "layers": 36,
      "features": [
        "vision"
      ],
      "mmlu_score": 65.6,
      "swe_bench_score": null,
      "notes": "Vision-language model. Same text backbone as Qwen 2.5 3B. 36 layers, 2 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "qwen-2.5-vl-72b-q4",
      "name": "Qwen 2.5 VL 72B",
      "url": "https://whatmodelscanirun.com/model/qwen-2-5-vl-72b",
      "quantization": "Q4_K_M",
      "params_b": 72.7,
      "weight_gb": 44.16,
      "gguf_repo": "bartowski/Qwen_Qwen2.5-VL-72B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.305,
      "max_context_k": 128,
      "layers": 80,
      "features": [
        "tool_use",
        "vision"
      ],
      "mmlu_score": 86.1,
      "swe_bench_score": null,
      "notes": "Vision-language model. Same text backbone as Qwen 2.5 72B. 80 layers, 8 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "qwen-2.5-vl-72b-q8",
      "name": "Qwen 2.5 VL 72B",
      "url": "https://whatmodelscanirun.com/model/qwen-2-5-vl-72b",
      "quantization": "Q8_0",
      "params_b": 72.7,
      "weight_gb": 71.96,
      "gguf_repo": "bartowski/Qwen_Qwen2.5-VL-72B-Instruct-GGUF",
      "gguf_shards": 2,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.305,
      "max_context_k": 128,
      "layers": 80,
      "features": [
        "tool_use",
        "vision"
      ],
      "mmlu_score": 86.1,
      "swe_bench_score": null,
      "notes": "Vision-language model. Same text backbone as Qwen 2.5 72B. 80 layers, 8 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "qwen-2.5-vl-7b-q4",
      "name": "Qwen 2.5 VL 7B",
      "url": "https://whatmodelscanirun.com/model/qwen-2-5-vl-7b",
      "quantization": "Q4_K_M",
      "params_b": 7.62,
      "weight_gb": 4.36,
      "gguf_repo": "unsloth/Qwen2.5-VL-7B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.053,
      "max_context_k": 128,
      "layers": 28,
      "features": [
        "tool_use",
        "vision"
      ],
      "mmlu_score": 74.2,
      "swe_bench_score": null,
      "notes": "Vision-language model. Same text backbone as Qwen 2.5 7B. 28 layers, 4 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "qwen-2.5-vl-7b-q8",
      "name": "Qwen 2.5 VL 7B",
      "url": "https://whatmodelscanirun.com/model/qwen-2-5-vl-7b",
      "quantization": "Q8_0",
      "params_b": 7.62,
      "weight_gb": 7.54,
      "gguf_repo": "unsloth/Qwen2.5-VL-7B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.053,
      "max_context_k": 128,
      "layers": 28,
      "features": [
        "tool_use",
        "vision"
      ],
      "mmlu_score": 74.2,
      "swe_bench_score": null,
      "notes": "Vision-language model. Same text backbone as Qwen 2.5 7B. 28 layers, 4 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "qwen3-32b-q4",
      "name": "Qwen3 32B",
      "url": "https://whatmodelscanirun.com/model/qwen3-32b",
      "quantization": "Q4_K_M",
      "params_b": 32.8,
      "weight_gb": 18.4,
      "gguf_repo": "Qwen/Qwen3-32B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.244,
      "max_context_k": 128,
      "layers": 64,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 85,
      "swe_bench_score": null,
      "notes": "Reasoning model (thinking/non-thinking modes). 64 layers, 8 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "qwen3-32b-q8",
      "name": "Qwen3 32B",
      "url": "https://whatmodelscanirun.com/model/qwen3-32b",
      "quantization": "Q8_0",
      "params_b": 32.8,
      "weight_gb": 32.43,
      "gguf_repo": "Qwen/Qwen3-32B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.244,
      "max_context_k": 128,
      "layers": 64,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 85,
      "swe_bench_score": null,
      "notes": "Reasoning model (thinking/non-thinking modes). 64 layers, 8 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "qwen3-8b-q4",
      "name": "Qwen3 8B",
      "url": "https://whatmodelscanirun.com/model/qwen3-8b",
      "quantization": "Q4_K_M",
      "params_b": 8.2,
      "weight_gb": 4.68,
      "gguf_repo": "Qwen/Qwen3-8B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.137,
      "max_context_k": 128,
      "layers": 36,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 79,
      "swe_bench_score": null,
      "notes": "Reasoning model (thinking/non-thinking modes). 36 layers, 8 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "qwen3-8b-q8",
      "name": "Qwen3 8B",
      "url": "https://whatmodelscanirun.com/model/qwen3-8b",
      "quantization": "Q8_0",
      "params_b": 8.2,
      "weight_gb": 8.11,
      "gguf_repo": "Qwen/Qwen3-8B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.137,
      "max_context_k": 128,
      "layers": 36,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 79,
      "swe_bench_score": null,
      "notes": "Reasoning model (thinking/non-thinking modes). 36 layers, 8 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "qwen3-coder-30b-a3b-q4",
      "name": "Qwen3 Coder 30B-A3B",
      "url": "https://whatmodelscanirun.com/model/qwen3-coder-30b-a3b",
      "quantization": "Q4_K_M",
      "params_b": 30.5,
      "weight_gb": 17.28,
      "gguf_repo": "unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 1.7,
      "kv_per_1k_gb": 0.092,
      "max_context_k": 256,
      "layers": 48,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 81,
      "swe_bench_score": null,
      "notes": "MoE: 30.5B total, 3.3B active (128 experts, 8 active). Coding-specialized, non-thinking mode only. 48 layers, 4 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "qwen3-coder-30b-a3b-q8",
      "name": "Qwen3 Coder 30B-A3B",
      "url": "https://whatmodelscanirun.com/model/qwen3-coder-30b-a3b",
      "quantization": "Q8_0",
      "params_b": 30.5,
      "weight_gb": 30.25,
      "gguf_repo": "unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 2.98,
      "kv_per_1k_gb": 0.092,
      "max_context_k": 256,
      "layers": 48,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 81,
      "swe_bench_score": null,
      "notes": "MoE: 30.5B total, 3.3B active (128 experts, 8 active). Coding-specialized, non-thinking mode only. 48 layers, 4 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "qwen3-next-80b-a3b-q4",
      "name": "Qwen3 Next 80B-A3B",
      "url": "https://whatmodelscanirun.com/model/qwen3-next-80b-a3b",
      "quantization": "Q4_K_M",
      "params_b": 80,
      "weight_gb": 45.09,
      "gguf_repo": "Qwen/Qwen3-Next-80B-A3B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 1.69,
      "kv_per_1k_gb": 0.023,
      "max_context_k": 256,
      "layers": 48,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 90,
      "swe_bench_score": null,
      "notes": "MoE: 80B total, 3B active (512 experts). Hybrid architecture (Gated DeltaNet + Gated Attention). Only 12 of 48 layers use attention (2 KV heads, head_dim 256). KV estimate covers attention layers only; DeltaNet layers use fixed-size recurrent state. Instruct-only (no thinking mode). GGUF Q4_K_M file size."
    },
    {
      "id": "qwen3-next-80b-a3b-q8",
      "name": "Qwen3 Next 80B-A3B",
      "url": "https://whatmodelscanirun.com/model/qwen3-next-80b-a3b",
      "quantization": "Q8_0",
      "params_b": 80,
      "weight_gb": 78.99,
      "gguf_repo": "Qwen/Qwen3-Next-80B-A3B-Instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 2.96,
      "kv_per_1k_gb": 0.023,
      "max_context_k": 256,
      "layers": 48,
      "features": [
        "tool_use"
      ],
      "mmlu_score": 90,
      "swe_bench_score": null,
      "notes": "MoE: 80B total, 3B active (512 experts). Hybrid architecture (Gated DeltaNet + Gated Attention). Only 12 of 48 layers use attention (2 KV heads, head_dim 256). KV estimate covers attention layers only; DeltaNet layers use fixed-size recurrent state. Instruct-only (no thinking mode). GGUF Q8_0 file size."
    },
    {
      "id": "qwen3.5-0.8b-q4",
      "name": "Qwen3.5 0.8B",
      "url": "https://whatmodelscanirun.com/model/qwen3-5-0-8b",
      "quantization": "Q4_K_M",
      "params_b": 0.8,
      "weight_gb": 0.5,
      "gguf_repo": "unsloth/Qwen3.5-0.8B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.011,
      "max_context_k": 256,
      "layers": 24,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 29.7,
      "swe_bench_score": null,
      "notes": "Dense multimodal model with hybrid attention: most layers are linear-attention with fixed-size recurrent state, so only the full-attention layers contribute a growing KV cache. Native 262K context. Apache 2.0. GGUF Q4_K_M file size."
    },
    {
      "id": "qwen3.5-0.8b-q8",
      "name": "Qwen3.5 0.8B",
      "url": "https://whatmodelscanirun.com/model/qwen3-5-0-8b",
      "quantization": "Q8_0",
      "params_b": 0.8,
      "weight_gb": 0.76,
      "gguf_repo": "unsloth/Qwen3.5-0.8B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.011,
      "max_context_k": 256,
      "layers": 24,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 29.7,
      "swe_bench_score": null,
      "notes": "Dense multimodal model with hybrid attention: most layers are linear-attention with fixed-size recurrent state, so only the full-attention layers contribute a growing KV cache. Native 262K context. Apache 2.0. GGUF Q8_0 file size."
    },
    {
      "id": "qwen3.5-122b-a10b-q4",
      "name": "Qwen3.5 122B A10B",
      "url": "https://whatmodelscanirun.com/model/qwen3-5-122b-a10b",
      "quantization": "Q4_K_M",
      "params_b": 122,
      "weight_gb": 71.28,
      "gguf_repo": "unsloth/Qwen3.5-122B-A10B-GGUF",
      "gguf_shards": 3,
      "active_weight_gb": 5.84,
      "kv_per_1k_gb": 0.023,
      "max_context_k": 256,
      "layers": 48,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 85,
      "swe_bench_score": 72,
      "notes": "Sparse MoE, 122B total with ~10B active per token. 48 layers, 12 full-attention (2 KV heads, head_dim 256), 36 linear-attention. Native 262K context. Multimodal (separate mmproj vision projector, not counted in the weight). SWE-bench Verified 72.0. GGUF Q4_K_M file size."
    },
    {
      "id": "qwen3.5-122b-a10b-q8",
      "name": "Qwen3.5 122B A10B",
      "url": "https://whatmodelscanirun.com/model/qwen3-5-122b-a10b",
      "quantization": "Q8_0",
      "params_b": 122,
      "weight_gb": 120.95,
      "gguf_repo": "unsloth/Qwen3.5-122B-A10B-GGUF",
      "gguf_shards": 4,
      "active_weight_gb": 9.91,
      "kv_per_1k_gb": 0.023,
      "max_context_k": 256,
      "layers": 48,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 85,
      "swe_bench_score": 72,
      "notes": "Sparse MoE, 122B total with ~10B active per token. 48 layers, 12 full-attention (2 KV heads, head_dim 256), 36 linear-attention. Native 262K context. Multimodal (separate mmproj vision projector, not counted in the weight). SWE-bench Verified 72.0. GGUF Q8_0 file size."
    },
    {
      "id": "qwen3.5-27b-q4",
      "name": "Qwen3.5 27B",
      "url": "https://whatmodelscanirun.com/model/qwen3-5-27b",
      "quantization": "Q4_K_M",
      "params_b": 27,
      "weight_gb": 15.59,
      "gguf_repo": "unsloth/Qwen3.5-27B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.061,
      "max_context_k": 256,
      "layers": 64,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 86.1,
      "swe_bench_score": 72.4,
      "notes": "Dense multimodal model with hybrid attention: most layers are linear-attention with fixed-size recurrent state, so only the full-attention layers contribute a growing KV cache. Native 262K context. Apache 2.0. GGUF Q4_K_M file size."
    },
    {
      "id": "qwen3.5-27b-q8",
      "name": "Qwen3.5 27B",
      "url": "https://whatmodelscanirun.com/model/qwen3-5-27b",
      "quantization": "Q8_0",
      "params_b": 27,
      "weight_gb": 26.63,
      "gguf_repo": "unsloth/Qwen3.5-27B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.061,
      "max_context_k": 256,
      "layers": 64,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 86.1,
      "swe_bench_score": 72.4,
      "notes": "Dense multimodal model with hybrid attention: most layers are linear-attention with fixed-size recurrent state, so only the full-attention layers contribute a growing KV cache. Native 262K context. Apache 2.0. GGUF Q8_0 file size."
    },
    {
      "id": "qwen3.5-2b-q4",
      "name": "Qwen3.5 2B",
      "url": "https://whatmodelscanirun.com/model/qwen3-5-2b",
      "quantization": "Q4_K_M",
      "params_b": 2,
      "weight_gb": 1.19,
      "gguf_repo": "unsloth/Qwen3.5-2B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.011,
      "max_context_k": 256,
      "layers": 24,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 55.3,
      "swe_bench_score": null,
      "notes": "Dense multimodal model with hybrid attention: most layers are linear-attention with fixed-size recurrent state, so only the full-attention layers contribute a growing KV cache. Native 262K context. Apache 2.0. GGUF Q4_K_M file size."
    },
    {
      "id": "qwen3.5-2b-q8",
      "name": "Qwen3.5 2B",
      "url": "https://whatmodelscanirun.com/model/qwen3-5-2b",
      "quantization": "Q8_0",
      "params_b": 2,
      "weight_gb": 1.87,
      "gguf_repo": "unsloth/Qwen3.5-2B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.011,
      "max_context_k": 256,
      "layers": 24,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 55.3,
      "swe_bench_score": null,
      "notes": "Dense multimodal model with hybrid attention: most layers are linear-attention with fixed-size recurrent state, so only the full-attention layers contribute a growing KV cache. Native 262K context. Apache 2.0. GGUF Q8_0 file size."
    },
    {
      "id": "qwen3.5-35b-a3b-q4",
      "name": "Qwen3.5 35B A3B",
      "url": "https://whatmodelscanirun.com/model/qwen3-5-35b-a3b",
      "quantization": "Q4_K_M",
      "params_b": 35,
      "weight_gb": 20.5,
      "gguf_repo": "unsloth/Qwen3.5-35B-A3B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 1.76,
      "kv_per_1k_gb": 0.019,
      "max_context_k": 256,
      "layers": 40,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 85.3,
      "swe_bench_score": 69.2,
      "notes": "Sparse MoE, 35B total with ~3B active per token. Hybrid attention: only the full-attention layers hold a growing KV cache. Native 262K context. Apache 2.0. GGUF Q4_K_M file size."
    },
    {
      "id": "qwen3.5-35b-a3b-q8",
      "name": "Qwen3.5 35B A3B",
      "url": "https://whatmodelscanirun.com/model/qwen3-5-35b-a3b",
      "quantization": "Q8_0",
      "params_b": 35,
      "weight_gb": 34.37,
      "gguf_repo": "unsloth/Qwen3.5-35B-A3B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 2.95,
      "kv_per_1k_gb": 0.019,
      "max_context_k": 256,
      "layers": 40,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 85.3,
      "swe_bench_score": 69.2,
      "notes": "Sparse MoE, 35B total with ~3B active per token. Hybrid attention: only the full-attention layers hold a growing KV cache. Native 262K context. Apache 2.0. GGUF Q8_0 file size."
    },
    {
      "id": "qwen3.5-397b-a17b-q4",
      "name": "Qwen3.5 397B A17B",
      "url": "https://whatmodelscanirun.com/model/qwen3-5-397b-a17b",
      "quantization": "Q4_K_M",
      "params_b": 397,
      "weight_gb": 227.33,
      "gguf_repo": "unsloth/Qwen3.5-397B-A17B-GGUF",
      "gguf_shards": 6,
      "active_weight_gb": 9.73,
      "kv_per_1k_gb": 0.029,
      "max_context_k": 256,
      "layers": 60,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 87,
      "swe_bench_score": 76.2,
      "notes": "Sparse MoE, 397B total with ~17B active per token. 60 layers, 15 full-attention (2 KV heads, head_dim 256), 45 linear-attention. Native 262K context. Multimodal (separate mmproj vision projector, not counted in the weight). SWE-bench Verified 76.2. GGUF Q4_K_M file size."
    },
    {
      "id": "qwen3.5-397b-a17b-q8",
      "name": "Qwen3.5 397B A17B",
      "url": "https://whatmodelscanirun.com/model/qwen3-5-397b-a17b",
      "quantization": "Q8_0",
      "params_b": 397,
      "weight_gb": 392.56,
      "gguf_repo": "unsloth/Qwen3.5-397B-A17B-GGUF",
      "gguf_shards": 10,
      "active_weight_gb": 16.81,
      "kv_per_1k_gb": 0.029,
      "max_context_k": 256,
      "layers": 60,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 87,
      "swe_bench_score": 76.2,
      "notes": "Sparse MoE, 397B total with ~17B active per token. 60 layers, 15 full-attention (2 KV heads, head_dim 256), 45 linear-attention. Native 262K context. Multimodal (separate mmproj vision projector, not counted in the weight). SWE-bench Verified 76.2. GGUF Q8_0 file size."
    },
    {
      "id": "qwen3.5-4b-q4",
      "name": "Qwen3.5 4B",
      "url": "https://whatmodelscanirun.com/model/qwen3-5-4b",
      "quantization": "Q4_K_M",
      "params_b": 4,
      "weight_gb": 2.55,
      "gguf_repo": "unsloth/Qwen3.5-4B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.031,
      "max_context_k": 256,
      "layers": 32,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 79.1,
      "swe_bench_score": null,
      "notes": "Dense multimodal model with hybrid attention: most layers are linear-attention with fixed-size recurrent state, so only the full-attention layers contribute a growing KV cache. Native 262K context. Apache 2.0. GGUF Q4_K_M file size."
    },
    {
      "id": "qwen3.5-4b-q8",
      "name": "Qwen3.5 4B",
      "url": "https://whatmodelscanirun.com/model/qwen3-5-4b",
      "quantization": "Q8_0",
      "params_b": 4,
      "weight_gb": 4.17,
      "gguf_repo": "unsloth/Qwen3.5-4B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.031,
      "max_context_k": 256,
      "layers": 32,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 79.1,
      "swe_bench_score": null,
      "notes": "Dense multimodal model with hybrid attention: most layers are linear-attention with fixed-size recurrent state, so only the full-attention layers contribute a growing KV cache. Native 262K context. Apache 2.0. GGUF Q8_0 file size."
    },
    {
      "id": "qwen3.5-9b-q4",
      "name": "Qwen3.5 9B",
      "url": "https://whatmodelscanirun.com/model/qwen3-5-9b",
      "quantization": "Q4_K_M",
      "params_b": 9,
      "weight_gb": 5.29,
      "gguf_repo": "unsloth/Qwen3.5-9B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.031,
      "max_context_k": 256,
      "layers": 32,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 82.5,
      "swe_bench_score": null,
      "notes": "Dense multimodal model with hybrid attention: most layers are linear-attention with fixed-size recurrent state, so only the full-attention layers contribute a growing KV cache. Native 262K context. Apache 2.0. GGUF Q4_K_M file size."
    },
    {
      "id": "qwen3.5-9b-q8",
      "name": "Qwen3.5 9B",
      "url": "https://whatmodelscanirun.com/model/qwen3-5-9b",
      "quantization": "Q8_0",
      "params_b": 9,
      "weight_gb": 8.87,
      "gguf_repo": "unsloth/Qwen3.5-9B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.031,
      "max_context_k": 256,
      "layers": 32,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 82.5,
      "swe_bench_score": null,
      "notes": "Dense multimodal model with hybrid attention: most layers are linear-attention with fixed-size recurrent state, so only the full-attention layers contribute a growing KV cache. Native 262K context. Apache 2.0. GGUF Q8_0 file size."
    },
    {
      "id": "qwen3.6-27b-q4",
      "name": "Qwen3.6 27B",
      "url": "https://whatmodelscanirun.com/model/qwen3-6-27b",
      "quantization": "Q4_K_M",
      "params_b": 27,
      "weight_gb": 15.41,
      "gguf_repo": "lmstudio-community/Qwen3.6-27B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.064,
      "max_context_k": 256,
      "layers": 64,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 86.2,
      "swe_bench_score": 77.2,
      "notes": "Dense multimodal model. Hybrid architecture: 16 blocks of (3 Gated DeltaNet -> FFN + 1 Gated Attention -> FFN). 64 total layers, only 16 use attention (24 Q heads, 4 KV heads, head_dim 256, RoPE rotation_dim 64). DeltaNet sublayers use 48 V heads, 16 QK heads, head_dim 128. KV estimate covers attention layers; DeltaNet layers use fixed-size recurrent state. Hidden dim 5120, FFN intermediate 17,408, vocab 248,320. Multi-Token Prediction enables speculative decoding. Native 262K context, extensible to ~1M tokens via YaRN. MMLU-Pro 86.2 (MMLU-Redux 93.5), SWE-bench Verified 77.2, Terminal-Bench 2.0 59.3, GPQA Diamond 87.8, AIME 2026 94.1, LiveCodeBench v6 83.9. Apache 2.0. GGUF Q4_K_M file size estimate."
    },
    {
      "id": "qwen3.6-27b-q8",
      "name": "Qwen3.6 27B",
      "url": "https://whatmodelscanirun.com/model/qwen3-6-27b",
      "quantization": "Q8_0",
      "params_b": 27,
      "weight_gb": 28,
      "gguf_repo": null,
      "gguf_shards": null,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.064,
      "max_context_k": 256,
      "layers": 64,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 86.2,
      "swe_bench_score": 77.2,
      "notes": "Dense multimodal model. Hybrid architecture: 16 blocks of (3 Gated DeltaNet -> FFN + 1 Gated Attention -> FFN). 64 total layers, only 16 use attention (24 Q heads, 4 KV heads, head_dim 256, RoPE rotation_dim 64). DeltaNet sublayers use 48 V heads, 16 QK heads, head_dim 128. KV estimate covers attention layers; DeltaNet layers use fixed-size recurrent state. Hidden dim 5120, FFN intermediate 17,408, vocab 248,320. Multi-Token Prediction enables speculative decoding. Native 262K context, extensible to ~1M tokens via YaRN. MMLU-Pro 86.2 (MMLU-Redux 93.5), SWE-bench Verified 77.2, Terminal-Bench 2.0 59.3, GPQA Diamond 87.8, AIME 2026 94.1, LiveCodeBench v6 83.9. Apache 2.0. GGUF Q8_0 file size estimate."
    },
    {
      "id": "qwen3.6-35b-a3b-q4",
      "name": "Qwen3.6 35B A3B",
      "url": "https://whatmodelscanirun.com/model/qwen3-6-35b-a3b",
      "quantization": "Q4_K_M",
      "params_b": 35,
      "weight_gb": 20.61,
      "gguf_repo": "unsloth/Qwen3.6-35B-A3B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 1.77,
      "kv_per_1k_gb": 0.01,
      "max_context_k": 256,
      "layers": 40,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 85.2,
      "swe_bench_score": 73.4,
      "notes": "Multimodal sparse MoE: 35B total, 3B active (256 experts, 8 routed + 1 shared per token). Hybrid architecture: 10 blocks of (3 Gated DeltaNet -> MoE + 1 Gated Attention -> MoE). 40 total layers, only 10 use attention (16 Q heads, 2 KV heads, head_dim 128). KV estimate covers attention layers; DeltaNet layers use fixed-size recurrent state. Native 262K context, extensible to ~1M tokens via YaRN. MMLU-Pro 85.2 (MMLU-Redux 93.3), SWE-bench Verified 73.4, AIME 2026 92.7, GPQA Diamond 86.0, MMMU 81.7. Apache 2.0. GGUF Q4_K_M file size."
    },
    {
      "id": "qwen3.6-35b-a3b-q8",
      "name": "Qwen3.6 35B A3B",
      "url": "https://whatmodelscanirun.com/model/qwen3-6-35b-a3b",
      "quantization": "Q8_0",
      "params_b": 35,
      "weight_gb": 34.37,
      "gguf_repo": "unsloth/Qwen3.6-35B-A3B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": 2.95,
      "kv_per_1k_gb": 0.01,
      "max_context_k": 256,
      "layers": 40,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 85.2,
      "swe_bench_score": 73.4,
      "notes": "Multimodal sparse MoE: 35B total, 3B active (256 experts, 8 routed + 1 shared per token). Hybrid architecture: 10 blocks of (3 Gated DeltaNet -> MoE + 1 Gated Attention -> MoE). 40 total layers, only 10 use attention (16 Q heads, 2 KV heads, head_dim 128). KV estimate covers attention layers; DeltaNet layers use fixed-size recurrent state. Native 262K context, extensible to ~1M tokens via YaRN. MMLU-Pro 85.2 (MMLU-Redux 93.3), SWE-bench Verified 73.4, AIME 2026 92.7, GPQA Diamond 86.0, MMMU 81.7. Apache 2.0. GGUF Q8_0 file size."
    },
    {
      "id": "qwen3.8-27b-q4",
      "name": "Qwen3.8 27B",
      "url": "https://whatmodelscanirun.com/model/qwen3-8-27b",
      "quantization": "Q4_K_M",
      "params_b": 27.78,
      "weight_gb": 15.33,
      "gguf_repo": "unsloth/Qwen3.8-27B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.061,
      "max_context_k": 256,
      "layers": 64,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 84.3,
      "swe_bench_score": 86,
      "notes": "Dense multimodal model. Hybrid architecture: 64 layers, only 16 use full attention (4 KV heads, head_dim 256); the other 48 are linear-attention layers with fixed-size recurrent state, so KV per 1K counts the attention layers only. Native 262K context, extensible to ~1M via YaRN. Accepts text, images and video. GPQA Diamond 89.2, LiveCodeBench v6 90.3, Terminal-Bench 2.1 73.0, OSWorld-Verified 84.3, SWE-bench Pro 61.7, IFBench 79.5. MMLU-Pro 84.3 and SWE-bench 86.0 are Vals AI third-party runs; Alibaba did not publish MMLU-Pro or SWE-bench Verified. Apache 2.0. GGUF Q4_K_M file size."
    },
    {
      "id": "qwen3.8-27b-q8",
      "name": "Qwen3.8 27B",
      "url": "https://whatmodelscanirun.com/model/qwen3-8-27b",
      "quantization": "Q8_0",
      "params_b": 27.78,
      "weight_gb": 27.05,
      "gguf_repo": "unsloth/Qwen3.8-27B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.061,
      "max_context_k": 256,
      "layers": 64,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 84.3,
      "swe_bench_score": 86,
      "notes": "Dense multimodal model. Hybrid architecture: 64 layers, only 16 use full attention (4 KV heads, head_dim 256); the other 48 are linear-attention layers with fixed-size recurrent state, so KV per 1K counts the attention layers only. Native 262K context, extensible to ~1M via YaRN. Accepts text, images and video. GPQA Diamond 89.2, LiveCodeBench v6 90.3, Terminal-Bench 2.1 73.0, OSWorld-Verified 84.3, SWE-bench Pro 61.7, IFBench 79.5. MMLU-Pro 84.3 and SWE-bench 86.0 are Vals AI third-party runs; Alibaba did not publish MMLU-Pro or SWE-bench Verified. Apache 2.0. GGUF Q8_0 file size."
    },
    {
      "id": "qwen3.8-flash-next-q4",
      "name": "Qwen3.8 Flash Next",
      "url": "https://whatmodelscanirun.com/model/qwen3-8-flash-next",
      "quantization": "Q4_K_M",
      "params_b": 180,
      "weight_gb": 111.39,
      "gguf_repo": "bartowski/Qwen3.8-Flash-Next-GGUF",
      "gguf_shards": 4,
      "active_weight_gb": 3.71,
      "kv_per_1k_gb": 0.023,
      "max_context_k": 256,
      "layers": 48,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 73.23,
      "swe_bench_score": null,
      "notes": "Preview of the Qwen4 architecture. ~180B total on disk: a 125B MoE backbone (512 experts, 10 routed plus 1 shared, ~6B active per token) with a 51B n-gram embedding table and a 4B MTP head. 48 layers, 12 full-attention (2 KV heads, head_dim 256), 36 linear-attention. 262K context. MMLU-Pro 73.23, SWE-bench Pro 62.5, LiveCodeBench v6 91.9, GPQA 91.7. Weight excludes the separate MTP and mmproj files. GGUF Q4_K_M file size."
    },
    {
      "id": "qwen3.8-flash-next-q8",
      "name": "Qwen3.8 Flash Next",
      "url": "https://whatmodelscanirun.com/model/qwen3-8-flash-next",
      "quantization": "Q8_0",
      "params_b": 180,
      "weight_gb": 175.33,
      "gguf_repo": "bartowski/Qwen3.8-Flash-Next-GGUF",
      "gguf_shards": 6,
      "active_weight_gb": 5.84,
      "kv_per_1k_gb": 0.023,
      "max_context_k": 256,
      "layers": 48,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": 73.23,
      "swe_bench_score": null,
      "notes": "Preview of the Qwen4 architecture. ~180B total on disk: a 125B MoE backbone (512 experts, 10 routed plus 1 shared, ~6B active per token) with a 51B n-gram embedding table and a 4B MTP head. 48 layers, 12 full-attention (2 KV heads, head_dim 256), 36 linear-attention. 262K context. MMLU-Pro 73.23, SWE-bench Pro 62.5, LiveCodeBench v6 91.9, GPQA 91.7. Weight excludes the separate MTP and mmproj files. GGUF Q8_0 file size."
    },
    {
      "id": "qwq-32b-q4",
      "name": "QwQ 32B",
      "url": "https://whatmodelscanirun.com/model/qwq-32b",
      "quantization": "Q4_K_M",
      "params_b": 32.5,
      "weight_gb": 18.49,
      "gguf_repo": "Qwen/QwQ-32B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.244,
      "max_context_k": 128,
      "layers": 64,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 83,
      "swe_bench_score": null,
      "notes": "Reasoning model. Based on Qwen 2.5 32B. 64 layers, 8 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "qwq-32b-q8",
      "name": "QwQ 32B",
      "url": "https://whatmodelscanirun.com/model/qwq-32b",
      "quantization": "Q8_0",
      "params_b": 32.5,
      "weight_gb": 32.43,
      "gguf_repo": "Qwen/QwQ-32B-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.244,
      "max_context_k": 128,
      "layers": 64,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 83,
      "swe_bench_score": null,
      "notes": "Reasoning model. Based on Qwen 2.5 32B. 64 layers, 8 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "smallthinker-3b-q4",
      "name": "SmallThinker 3B",
      "url": "https://whatmodelscanirun.com/model/smallthinker-3b",
      "quantization": "Q4_K_M",
      "params_b": 3.09,
      "weight_gb": 1.96,
      "gguf_repo": "bartowski/SmallThinker-3B-Preview-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.034,
      "max_context_k": 32,
      "layers": 36,
      "features": [
        "reasoning"
      ],
      "mmlu_score": 65,
      "swe_bench_score": null,
      "notes": "Reasoning model. Fine-tuned from Qwen 2.5 3B Instruct. 36 layers, 2 KV heads, head_dim 128. GGUF Q4_K_M file size."
    },
    {
      "id": "smallthinker-3b-q8",
      "name": "SmallThinker 3B",
      "url": "https://whatmodelscanirun.com/model/smallthinker-3b",
      "quantization": "Q8_0",
      "params_b": 3.09,
      "weight_gb": 3.29,
      "gguf_repo": null,
      "gguf_shards": null,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.034,
      "max_context_k": 32,
      "layers": 36,
      "features": [
        "reasoning"
      ],
      "mmlu_score": 65,
      "swe_bench_score": null,
      "notes": "Reasoning model. Fine-tuned from Qwen 2.5 3B Instruct. 36 layers, 2 KV heads, head_dim 128. GGUF Q8_0 file size."
    },
    {
      "id": "starcoder2-15b-q4",
      "name": "StarCoder2 15B",
      "url": "https://whatmodelscanirun.com/model/starcoder2-15b",
      "quantization": "Q4_K_M",
      "params_b": 15.6,
      "weight_gb": 9.18,
      "gguf_repo": "bartowski/starcoder2-15b-instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.076,
      "max_context_k": 16,
      "layers": 40,
      "features": [],
      "mmlu_score": 51.4,
      "swe_bench_score": null,
      "notes": "Code model. BigCode. 40 layers, 4 KV heads, head_dim 128. Trained on 600+ languages. GGUF Q4_K_M file size."
    },
    {
      "id": "starcoder2-15b-q8",
      "name": "StarCoder2 15B",
      "url": "https://whatmodelscanirun.com/model/starcoder2-15b",
      "quantization": "Q8_0",
      "params_b": 15.6,
      "weight_gb": 15.8,
      "gguf_repo": "bartowski/starcoder2-15b-instruct-GGUF",
      "gguf_shards": 1,
      "active_weight_gb": null,
      "kv_per_1k_gb": 0.076,
      "max_context_k": 16,
      "layers": 40,
      "features": [],
      "mmlu_score": 51.4,
      "swe_bench_score": null,
      "notes": "Code model. BigCode. 40 layers, 4 KV heads, head_dim 128. Trained on 600+ languages. GGUF Q8_0 file size."
    },
    {
      "id": "step-3.5-flash-q4",
      "name": "Step 3.5 Flash",
      "url": "https://whatmodelscanirun.com/model/step-3-5-flash",
      "quantization": "Q4_K_M",
      "params_b": 197,
      "weight_gb": 111.7,
      "gguf_repo": "bartowski/stepfun-ai_Step-3.5-Flash-GGUF",
      "gguf_shards": 4,
      "active_weight_gb": 6.24,
      "kv_per_1k_gb": 0.172,
      "max_context_k": 256,
      "layers": 45,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 84,
      "swe_bench_score": 74.4,
      "notes": "MoE: 197B total, 11B active (288 experts, Top-8). 45 layers, 3:1 SWA ratio. KV estimate based on estimated 8 KV heads, head_dim 128. GGUF Q4_K_M file size. Active params: 11B (Step-3.5-Flash model card)."
    },
    {
      "id": "step-3.5-flash-q8",
      "name": "Step 3.5 Flash",
      "url": "https://whatmodelscanirun.com/model/step-3-5-flash",
      "quantization": "Q8_0",
      "params_b": 197,
      "weight_gb": 195.04,
      "gguf_repo": "bartowski/stepfun-ai_Step-3.5-Flash-GGUF",
      "gguf_shards": 6,
      "active_weight_gb": 10.89,
      "kv_per_1k_gb": 0.172,
      "max_context_k": 256,
      "layers": 45,
      "features": [
        "reasoning",
        "tool_use"
      ],
      "mmlu_score": 84,
      "swe_bench_score": 74.4,
      "notes": "MoE: 197B total, 11B active (288 experts, Top-8). 45 layers, 3:1 SWA ratio. KV estimate based on estimated 8 KV heads, head_dim 128. GGUF Q8_0 file size (estimated). Active params: 11B (Step-3.5-Flash model card)."
    },
    {
      "id": "step-3.7-flash-q4",
      "name": "Step 3.7 Flash",
      "url": "https://whatmodelscanirun.com/model/step-3-7-flash",
      "quantization": "Q4_K_S",
      "params_b": 198,
      "weight_gb": 103.84,
      "gguf_repo": "stepfun-ai/Step-3.7-Flash-GGUF",
      "gguf_shards": 3,
      "active_weight_gb": 5.77,
      "kv_per_1k_gb": 0.172,
      "max_context_k": 256,
      "layers": 45,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": null,
      "swe_bench_score": 76.5,
      "notes": "45 layers, 8 KV heads, head_dim 128; MoE 288 experts 8 active; vision encoder 1.8B Active params: 11B (Step-3.7-Flash model card)."
    },
    {
      "id": "step-3.7-flash-q8",
      "name": "Step 3.7 Flash",
      "url": "https://whatmodelscanirun.com/model/step-3-7-flash",
      "quantization": "Q8_0",
      "params_b": 198,
      "weight_gb": 195.04,
      "gguf_repo": "stepfun-ai/Step-3.7-Flash-GGUF",
      "gguf_shards": 5,
      "active_weight_gb": 10.84,
      "kv_per_1k_gb": 0.172,
      "max_context_k": 256,
      "layers": 45,
      "features": [
        "reasoning",
        "tool_use",
        "vision"
      ],
      "mmlu_score": null,
      "swe_bench_score": 76.5,
      "notes": "45 layers, 8 KV heads, head_dim 128; MoE 288 experts 8 active; vision encoder 1.8B Active params: 11B (Step-3.7-Flash model card)."
    }
  ]
}
