What Models?

Pick the right model for your GPU — in seconds.

What AI models can an M5 Ultra (96–512 GB) run?

The M5 Ultra has 512 GB of unified memory and 1200 GB/s memory bandwidth (speeds below assume 70% of it, the share llama.cpp reaches on dual-die Ultra chips). 101 models run well on it and 2 more are a tight fit. Best models for 128 GB →

Memory

As an Amazon Associate, I earn from qualifying purchases.

Best for general use

Highest MMLU score among models that run well

  1. DeepSeek R1 0528 Q4_K_M 160K context · ~33 tok/s · MMLU 90.8
  2. Qwen3 Next 80B-A3B Q8_0 256K context · ~134 tok/s · MMLU 90
  3. Mistral Small 4 Q8_0 718K context · ~84 tok/s · MMLU 88.5

Best for coding

Highest SWE-bench Verified score among models that run well

  1. Ornith 1.5 397B Q8_0 256K context · ~42 tok/s · SWE-bench 86
  2. Qwen3.8 27B Q8_0 256K context · ~28 tok/s · SWE-bench 86
  3. Ornith 1.5 122B Q8_0 256K context · ~61 tok/s · SWE-bench 84.5

Every model on the M5 Ultra (512 GB)

201 of 214 model configurations can run on 512 GB — sorted by quality

107 models, each counted once by the best way it runs on your hardware.

Largest that runs well Muse Spark Open 816B A42B 822.04B params · Q4_K_M · 765K context
Fastest good-quality model Phi-3 Mini 3.8B ~247 tok/s · Q4_K_M · 128K context
? MMLU (Massive Multitask Language Understanding) measures general knowledge across 57 subjects. Higher = more capable. Sorted best-first.

Runs well 101

  • DeepSeek R1 0528 671B Excellent Reasoning
    MMLU 90.8
    Q4_K_M
    Context 160K Speed ~33 tok/s

    MoE — speed based on ~21.02 GB active experts plus per-token routing overhead (full 381.12 GB loaded in VRAM)

    Q8_0 Doesn't fit · needs 664.4 GB
  • Qwen3 Next 80B-A3B 80B Excellent Tool use
    MMLU 90
    Q4_K_M
    Context 256K Speed ~171 tok/s

    MoE — speed based on ~1.69 GB active experts plus per-token routing overhead (full 45.09 GB loaded in VRAM)

    Q8_0
    Context 256K Speed ~134 tok/s

    MoE — speed based on ~2.96 GB active experts plus per-token routing overhead (full 78.99 GB loaded in VRAM)

  • Mistral Small 4 119B Excellent VisionTool use
    MMLU 88.5
    Q4_K_M
    Context 809K Speed ~119 tok/s

    MoE — speed based on ~3.7 GB active experts plus per-token routing overhead (full 67.65 GB loaded in VRAM)

    Q8_0
    Context 718K Speed ~84 tok/s

    MoE — speed based on ~6.43 GB active experts plus per-token routing overhead (full 117.79 GB loaded in VRAM)

Runs well 101

  • DeepSeek R1 0528 671B Excellent Reasoning
    MMLU 90.8
    Q4_K_M
    Context 160K Speed ~33 tok/s

    MoE — speed based on ~21.02 GB active experts plus per-token routing overhead (full 381.12 GB loaded in VRAM)

    Q8_0 Doesn't fit · needs 664.4 GB
  • Qwen3 Next 80B-A3B 80B Excellent Tool use
    MMLU 90
    Q4_K_M
    Context 256K Speed ~171 tok/s

    MoE — speed based on ~1.69 GB active experts plus per-token routing overhead (full 45.09 GB loaded in VRAM)

    Q8_0
    Context 256K Speed ~134 tok/s

    MoE — speed based on ~2.96 GB active experts plus per-token routing overhead (full 78.99 GB loaded in VRAM)

  • Mistral Small 4 119B Excellent VisionTool use
    MMLU 88.5
    Q4_K_M
    Context 809K Speed ~119 tok/s

    MoE — speed based on ~3.7 GB active experts plus per-token routing overhead (full 67.65 GB loaded in VRAM)

    Q8_0
    Context 718K Speed ~84 tok/s

    MoE — speed based on ~6.43 GB active experts plus per-token routing overhead (full 117.79 GB loaded in VRAM)

  • MiniMax M2.1 229B Excellent ReasoningTool use
    MMLU 88
    Q4_K_M
    Context 128K Speed ~92 tok/s

    MoE — speed based on ~5.63 GB active experts plus per-token routing overhead (full 128.84 GB loaded in VRAM)

    Q8_0
    Context 128K Speed ~61 tok/s

    MoE — speed based on ~9.89 GB active experts plus per-token routing overhead (full 226.44 GB loaded in VRAM)

  • Qwen3.5 397B A17B 397B Excellent VisionTool useReasoning
    MMLU 87
    Q4_K_M
    Context 256K Speed ~62 tok/s

    MoE — speed based on ~9.73 GB active experts plus per-token routing overhead (full 227.33 GB loaded in VRAM)

    Q8_0
    Context 256K Speed ~40 tok/s

    MoE — speed based on ~16.81 GB active experts plus per-token routing overhead (full 392.56 GB loaded in VRAM)

  • GLM-4.6 357B Excellent ReasoningTool use
    MMLU 87
    Q4_K_M
    Context 200K Speed ~37 tok/s

    MoE — speed based on ~18 GB active experts plus per-token routing overhead (full 200.78 GB loaded in VRAM)

    Q8_0
    Context 200K Speed ~23 tok/s

    MoE — speed based on ~31.67 GB active experts plus per-token routing overhead (full 353.27 GB loaded in VRAM)

  • GLM 5.1 754B Excellent Tool use
    MMLU 86.9
    Q4_K_M
    Context 198K Speed ~31 tok/s

    MoE — speed based on ~22.71 GB active experts plus per-token routing overhead (full 428.11 GB loaded in VRAM)

    Q8_0 Doesn't fit · needs 746.4 GB
  • DeepSeek V4 Flash 284B Excellent Tool useReasoning
    MMLU 86.4
    Q4_K_M
    Context 1M Speed ~76 tok/s

    MoE — speed based on ~7.37 GB active experts plus per-token routing overhead (full 161 GB loaded in VRAM)

    Q8_0
    Context 1M Speed ~50 tok/s

    MoE — speed based on ~12.91 GB active experts plus per-token routing overhead (full 282 GB loaded in VRAM)

  • Qwen3.6 27B 27B Excellent VisionTool useReasoning
    MMLU 86.2
    Q4_K_M
    Context 256K Speed ~48 tok/s
    Q8_0
    Context 256K Speed ~27 tok/s
  • GLM 5.3 Flash 321.32B Excellent Tool useReasoning
    MMLU 86.1
    Q4_K_M
    Context 1M Speed ~61 tok/s

    MoE — speed based on ~9.86 GB active experts plus per-token routing overhead (full 176.02 GB loaded in VRAM)

    Q8_0
    Context 1M Speed ~39 tok/s

    MoE — speed based on ~17.39 GB active experts plus per-token routing overhead (full 310.41 GB loaded in VRAM)

  • Qwen3.5 27B 27B Excellent VisionTool useReasoning
    MMLU 86.1
    Q4_K_M
    Context 256K Speed ~47 tok/s
    Q8_0
    Context 256K Speed ~28 tok/s
  • Qwen 2.5 72B 72.7B Excellent Tool use
    MMLU 86.1
    Q4_K_M
    Context 128K Speed ~17 tok/s
    Q8_0
    Context 128K Speed ~11 tok/s
  • Qwen 2.5 VL 72B 72.7B Excellent VisionTool use
    MMLU 86.1
    Q4_K_M
    Context 128K Speed ~17 tok/s
    Q8_0
    Context 128K Speed ~11 tok/s
  • Llama 3.3 70B 70.6B Excellent Tool use
    MMLU 86
    Q4_K_M
    Context 128K Speed ~19 tok/s
    Q8_0
    Context 128K Speed ~11 tok/s
  • Llama 4 Maverick 400B Excellent VisionTool use
    MMLU 85.5
    Q4_K_M
    Context 128K Speed ~63 tok/s

    MoE — speed based on ~9.61 GB active experts plus per-token routing overhead (full 226.09 GB loaded in VRAM)

    Q8_0
    Context 128K Speed ~40 tok/s

    MoE — speed based on ~16.85 GB active experts plus per-token routing overhead (full 396.57 GB loaded in VRAM)

  • Qwen3.5 35B A3B 35B Excellent VisionTool useReasoning
    MMLU 85.3
    Q4_K_M
    Context 256K Speed ~168 tok/s

    MoE — speed based on ~1.76 GB active experts plus per-token routing overhead (full 20.5 GB loaded in VRAM)

    Q8_0
    Context 256K Speed ~134 tok/s

    MoE — speed based on ~2.95 GB active experts plus per-token routing overhead (full 34.37 GB loaded in VRAM)

  • Gemma 4 31B 31B Excellent VisionTool useReasoning
    MMLU 85.2
    Q4_K_M
    Context 256K Speed ~43 tok/s
    Q8_0
    Context 256K Speed ~25 tok/s
  • Qwen3.6 35B A3B 35B Excellent VisionTool useReasoning
    MMLU 85.2
    Q4_K_M
    Context 256K Speed ~168 tok/s

    MoE — speed based on ~1.77 GB active experts plus per-token routing overhead (full 20.61 GB loaded in VRAM)

    Q8_0
    Context 256K Speed ~134 tok/s

    MoE — speed based on ~2.95 GB active experts plus per-token routing overhead (full 34.37 GB loaded in VRAM)

  • Qwen3.5 122B A10B 122B Excellent VisionTool useReasoning
    MMLU 85
    Q4_K_M
    Context 256K Speed ~90 tok/s

    MoE — speed based on ~5.84 GB active experts plus per-token routing overhead (full 71.28 GB loaded in VRAM)

    Q8_0
    Context 256K Speed ~61 tok/s

    MoE — speed based on ~9.91 GB active experts plus per-token routing overhead (full 120.95 GB loaded in VRAM)

  • Qwen3 32B 32.8B Excellent ReasoningTool use
    MMLU 85
    Q4_K_M
    Context 128K Speed ~40 tok/s
    Q8_0
    Context 128K Speed ~23 tok/s
  • Phi-4 14B 14.7B Excellent
    MMLU 84.8
    Q4_K_M
    Context 16K Speed ~84 tok/s
    Q8_0
    Context 16K Speed ~51 tok/s
  • Qwen3.8 27B 27.78B Excellent VisionTool useReasoning
    MMLU 84.3
    Q4_K_M
    Context 256K Speed ~48 tok/s
    Q8_0
    Context 256K Speed ~28 tok/s
  • Step 3.5 Flash 197B Excellent ReasoningTool use
    MMLU 84
    Q4_K_M
    Context 256K Speed ~86 tok/s

    MoE — speed based on ~6.24 GB active experts plus per-token routing overhead (full 111.7 GB loaded in VRAM)

    Q8_0
    Context 256K Speed ~57 tok/s

    MoE — speed based on ~10.89 GB active experts plus per-token routing overhead (full 195.04 GB loaded in VRAM)

  • DeepSeek R1 Distill Llama 70B 70.6B Excellent Reasoning
    MMLU 84
    Q4_K_M
    Context 128K Speed ~19 tok/s
    Q8_0
    Context 128K Speed ~11 tok/s
  • Qwen 2.5 32B 32.5B Excellent Tool use
    MMLU 83.3
    Q4_K_M
    Context 128K Speed ~40 tok/s
    Q8_0
    Context 128K Speed ~23 tok/s
  • GLM 5 753.86B Excellent Tool use
    MMLU 83
    Q4_K_M
    Context 198K Speed ~31 tok/s

    MoE — speed based on ~22.53 GB active experts plus per-token routing overhead (full 424.57 GB loaded in VRAM)

    Q8_0 Doesn't fit · needs 746.4 GB
  • QwQ 32B 32.5B Excellent ReasoningTool use
    MMLU 83
    Q4_K_M
    Context 128K Speed ~40 tok/s
    Q8_0
    Context 128K Speed ~23 tok/s
  • MiMo V2.5 310B Great Reasoning
    MMLU 82.9
    Q4_K_M
    Context 1M Speed ~69 tok/s

    MoE — speed based on ~8.51 GB active experts plus per-token routing overhead (full 175.8 GB loaded in VRAM)

    Q8_0
    Context 1M Speed ~44 tok/s

    MoE — speed based on ~14.84 GB active experts plus per-token routing overhead (full 306.67 GB loaded in VRAM)

  • Gemma 4 26B A4B 26B Great VisionTool useReasoning
    MMLU 82.6
    Q4_K_M
    Context 256K Speed ~147 tok/s

    MoE — speed based on ~2.44 GB active experts plus per-token routing overhead (full 15.87 GB loaded in VRAM)

    Q8_0
    Context 256K Speed ~116 tok/s

    MoE — speed based on ~3.85 GB active experts plus per-token routing overhead (full 25.02 GB loaded in VRAM)

  • Qwen3.5 9B 9B Great VisionTool useReasoning
    MMLU 82.5
    Q4_K_M
    Context 256K Speed ~126 tok/s
    Q8_0
    Context 256K Speed ~80 tok/s
  • MiniMax M2.5 228.7B Great ReasoningTool use
    MMLU 82
    Q4_K_M
    Context 192K Speed ~92 tok/s

    MoE — speed based on ~5.63 GB active experts plus per-token routing overhead (full 128.84 GB loaded in VRAM)

    Q8_0
    Context 192K Speed ~61 tok/s

    MoE — speed based on ~9.9 GB active experts plus per-token routing overhead (full 226.44 GB loaded in VRAM)

  • GLM-4.7 Flash 31B Great ReasoningTool use
    MMLU 82
    Q4_K_M
    Context 128K Speed ~172 tok/s

    MoE — speed based on ~1.65 GB active experts plus per-token routing overhead (full 17.05 GB loaded in VRAM)

    Q8_0
    Context 128K Speed ~136 tok/s

    MoE — speed based on ~2.87 GB active experts plus per-token routing overhead (full 29.66 GB loaded in VRAM)

  • Magistral Small 24B 23.6B Great ReasoningVisionTool use
    MMLU 82
    Q4_K_M
    Context 128K Speed ~55 tok/s
    Q8_0
    Context 128K Speed ~32 tok/s
  • Nemotron 3.5 Lightning 30B A3B 31.58B Great Tool useReasoning
    MMLU 81.94
    Q4_K_M
    Context 256K Speed ~152 tok/s

    MoE — speed based on ~2.25 GB active experts plus per-token routing overhead (full 23.73 GB loaded in VRAM)

    Q8_0
    Context 256K Speed ~131 tok/s

    MoE — speed based on ~3.1 GB active experts plus per-token routing overhead (full 32.6 GB loaded in VRAM)

  • Qwen3 Coder 30B-A3B 30.5B Great Tool use
    MMLU 81
    Q4_K_M
    Context 256K Speed ~171 tok/s

    MoE — speed based on ~1.7 GB active experts plus per-token routing overhead (full 17.28 GB loaded in VRAM)

    Q8_0
    Context 256K Speed ~133 tok/s

    MoE — speed based on ~2.98 GB active experts plus per-token routing overhead (full 30.25 GB loaded in VRAM)

  • Mistral Small 3.1 24B 23.6B Great VisionTool use
    MMLU 81
    Q4_K_M
    Context 128K Speed ~55 tok/s
    Q8_0
    Context 128K Speed ~32 tok/s
  • DeepSeek R1 Distill Qwen 32B 32.5B Great Reasoning
    MMLU 80
    Q4_K_M
    Context 128K Speed ~40 tok/s
    Q8_0
    Context 128K Speed ~23 tok/s
  • Qwen 2.5 14B 14.7B Great Tool use
    MMLU 79.9
    Q4_K_M
    Context 128K Speed ~84 tok/s
    Q8_0
    Context 128K Speed ~50 tok/s
  • Nemotron Cascade 2 30B A3B 30B Great ReasoningTool use
    MMLU 79.8
    Q4_K_M
    Context 256K Speed ~151 tok/s

    MoE — speed based on ~2.3 GB active experts plus per-token routing overhead (full 23.03 GB loaded in VRAM)

    Q8_0
    Context 256K Speed ~130 tok/s

    MoE — speed based on ~3.13 GB active experts plus per-token routing overhead (full 31.28 GB loaded in VRAM)

  • Llama 4 Scout 109B Great VisionTool use
    MMLU 79.6
    Q4_K_M
    Context 128K Speed ~63 tok/s

    MoE — speed based on ~9.49 GB active experts plus per-token routing overhead (full 60.87 GB loaded in VRAM)

    Q8_0
    Context 128K Speed ~40 tok/s

    MoE — speed based on ~16.64 GB active experts plus per-token routing overhead (full 106.67 GB loaded in VRAM)

  • Qwen3.5 4B 4B Great VisionTool useReasoning
    MMLU 79.1
    Q4_K_M
    Context 256K Speed ~225 tok/s
    Q8_0
    Context 256K Speed ~153 tok/s
  • Qwen3 8B 8.2B Great ReasoningTool use
    MMLU 79
    Q4_K_M
    Context 128K Speed ~139 tok/s
    Q8_0
    Context 128K Speed ~87 tok/s
  • Qwen 2.5 Coder 32B 32.5B Great
    MMLU 79
    Q4_K_M
    Context 128K Speed ~40 tok/s
    Q8_0
    Context 128K Speed ~23 tok/s
  • Gemma 3 27B 27B Great Vision
    MMLU 78
    Q4_K_M
    Context 128K Speed ~48 tok/s
    Q8_0
    Context 128K Speed ~28 tok/s
  • Gemma 4 12B 11.96B Great VisionTool useReasoning
    MMLU 77.2
    Q4_K_M
    Context 256K Speed ~103 tok/s
    Q8_0
    Context 256K Speed ~61 tok/s
  • Nemotron Nano 12B 12.6B Great ReasoningTool use
    MMLU 76
    Q4_K_M
    Context 128K Speed ~99 tok/s
    Q8_0
    Context 128K Speed ~60 tok/s
  • GPT-OSS 20B 21B Great ReasoningTool use
    MMLU 75
    Q4_K_M
    Context 128K Speed ~165 tok/s

    MoE — speed based on ~1.86 GB active experts plus per-token routing overhead (full 10.83 GB loaded in VRAM)

    Q8_0
    Context 128K Speed ~162 tok/s

    MoE — speed based on ~1.93 GB active experts plus per-token routing overhead (full 11.28 GB loaded in VRAM)

  • Qwen 2.5 7B 7.62B Good Tool use
    MMLU 74.2
    Q4_K_M
    Context 128K Speed ~148 tok/s
    Q8_0
    Context 128K Speed ~92 tok/s
  • Qwen 2.5 VL 7B 7.62B Good VisionTool use
    MMLU 74.2
    Q4_K_M
    Context 128K Speed ~148 tok/s
    Q8_0
    Context 128K Speed ~92 tok/s
  • Gemma 3 12B 11.8B Good Vision
    MMLU 74
    Q4_K_M
    Context 128K Speed ~101 tok/s
    Q8_0
    Context 128K Speed ~62 tok/s
  • Qwen 2.5 Coder 14B 14.7B Good
    MMLU 74
    Q4_K_M
    Context 128K Speed ~84 tok/s
    Q8_0
    Context 128K Speed ~50 tok/s
  • Qwen3.8 Flash Next 180B Good VisionTool useReasoning
    MMLU 73.23
    Q4_K_M
    Context 256K Speed ~119 tok/s

    MoE — speed based on ~3.71 GB active experts plus per-token routing overhead (full 111.39 GB loaded in VRAM)

    Q8_0
    Context 256K Speed ~90 tok/s

    MoE — speed based on ~5.84 GB active experts plus per-token routing overhead (full 175.33 GB loaded in VRAM)

  • Mistral Large 3 675B 673.42B Good VisionTool useReasoning
    MMLU 73.11
    Q4_K_M
    Context 288K Speed ~30 tok/s

    MoE — speed based on ~23.08 GB active experts plus per-token routing overhead (full 379.04 GB loaded in VRAM)

    Q8_0 Doesn't fit · needs 666.6 GB
  • Llama 3.2 Vision 11B 10.6B Good VisionTool use
    MMLU 73
    Q4_K_M
    Context 128K Speed ~114 tok/s
    Q8_0
    Context 128K Speed ~69 tok/s
  • DeepSeek R1 Distill Qwen 14B 14.7B Good Reasoning
    MMLU 72
    Q4_K_M
    Context 128K Speed ~84 tok/s
    Q8_0
    Context 128K Speed ~50 tok/s
  • Devstral Small 24B 23.6B Good Tool use
    MMLU 72
    Q4_K_M
    Context 128K Speed ~55 tok/s
    Q8_0
    Context 128K Speed ~32 tok/s
  • Phi-3 Mini 3.8B 3.82B Good
    MMLU 70.9
    Q4_K_M
    Context 128K Speed ~247 tok/s
    Q8_0
    Context 128K Speed ~166 tok/s
  • Phi-4 Mini 3.8B 3.84B Good
    MMLU 70
    Q4_K_M
    Context 128K Speed ~241 tok/s
    Q8_0
    Context 128K Speed ~165 tok/s
  • Llama 3.1 8B 8.03B Good Tool use
    MMLU 69.4
    Q4_K_M
    Context 128K Speed ~142 tok/s
    Q8_0
    Context 128K Speed ~88 tok/s
    fp16
    Context 128K Speed ~49 tok/s
  • Gemma 4 E4B 4.5B Good VisionTool useReasoning
    MMLU 69.4
    Q4_K_M
    Context 128K Speed ~131 tok/s
    Q8_0
    Context 128K Speed ~93 tok/s
  • Mistral Nemo 12B 12.2B Good Tool use
    MMLU 68
    Q4_K_M
    Context 128K Speed ~99 tok/s
    Q8_0
    Context 128K Speed ~60 tok/s
  • Qwen 2.5 Coder 7B 7.62B Good
    MMLU 68
    Q4_K_M
    Context 128K Speed ~148 tok/s
    Q8_0
    Context 128K Speed ~92 tok/s
  • Qwen 2.5 VL 3B 3.09B Fair Vision
    MMLU 65.6
    Q4_K_M
    Context 128K Speed ~286 tok/s
    Q8_0
    Context 128K Speed ~196 tok/s
  • Qwen 2.5 3B 3.09B Fair Tool use
    MMLU 65.6
    Q4_K_M
    Context 32K Speed ~286 tok/s
    Q8_0
    Context 32K Speed ~196 tok/s
  • Ministral 8B 8.02B Fair Tool use
    MMLU 65
    Q4_K_M
    Context 128K Speed ~130 tok/s
    Q8_0
    Context 128K Speed ~88 tok/s
  • SmallThinker 3B 3.09B Fair Reasoning
    MMLU 65
    Q4_K_M
    Context 32K Speed ~271 tok/s
    Q8_0
    Context 32K Speed ~185 tok/s
  • Llama 3.2 3B 3.21B Fair Tool use
    MMLU 63.4
    Q4_K_M
    Context 128K Speed ~278 tok/s
    Q8_0
    Context 128K Speed ~190 tok/s
    fp16
    Context 128K Speed ~113 tok/s
  • Gemma 3 4B 3.88B Fair Vision
    MMLU 63
    Q4_K_M
    Context 128K Speed ~241 tok/s
    Q8_0
    Context 128K Speed ~164 tok/s
  • Codestral 22B 22.2B Fair
    MMLU 62
    Q4_K_M
    Context 32K Speed ~59 tok/s
    Q8_0
    Context 32K Speed ~34 tok/s
  • Gemma 4 E2B 2.3B Fair VisionTool useReasoning
    MMLU 60
    Q4_K_M
    Context 128K Speed ~188 tok/s
    Q8_0
    Context 128K Speed ~141 tok/s
  • DeepSeek R1 Distill Qwen 7B 7.62B Fair Reasoning
    MMLU 58
    Q4_K_M
    Context 128K Speed ~148 tok/s
    Q8_0
    Context 128K Speed ~92 tok/s
  • Qwen3.5 2B 2B Fair VisionTool useReasoning
    MMLU 55.3
    Q4_K_M
    Context 256K Speed ~369 tok/s
    Q8_0
    Context 256K Speed ~279 tok/s
  • StarCoder2 15B 15.6B Basic
    MMLU 51.4
    Q4_K_M
    Context 16K Speed ~77 tok/s
    Q8_0
    Context 16K Speed ~47 tok/s
  • DeepSeek Coder 6.7B 6.74B Basic
    MMLU 50
    Q4_K_M
    Context 16K Speed ~165 tok/s
    Q8_0
    Context 16K Speed ~103 tok/s
  • Llama 3.2 1B 1.24B Basic Tool use
    MMLU 49.3
    Q4_K_M
    Context 128K Speed ~465 tok/s
    Q8_0
    Context 128K Speed ~362 tok/s
  • Qwen3.5 0.8B 0.8B Basic VisionTool useReasoning
    MMLU 29.7
    Q4_K_M
    Context 256K Speed ~547 tok/s
    Q8_0
    Context 256K Speed ~463 tok/s
  • Nemotron 3.5 Super VL 120B A12B 124.45B Basic VisionTool useReasoning
    MMLU
    Q4_K_M
    Context 1M Speed ~85 tok/s

    MoE — speed based on ~6.3 GB active experts plus per-token routing overhead (full 65.29 GB loaded in VRAM)

    Q8_0
    Context 1M Speed ~54 tok/s

    MoE — speed based on ~11.54 GB active experts plus per-token routing overhead (full 119.65 GB loaded in VRAM)

  • Beam 1.0 501B A23B 501.27B Basic Tool useReasoning
    MMLU
    Q4_K_M
    Context 1M Speed ~47 tok/s

    MoE — speed based on ~13.67 GB active experts plus per-token routing overhead (full 292.93 GB loaded in VRAM)

    Q8_0
    Context 464K Speed ~30 tok/s

    MoE — speed based on ~23.16 GB active experts plus per-token routing overhead (full 496.2 GB loaded in VRAM)

  • Muse Spark Open 816B A42B 822.04B Basic VisionTool useReasoning
    MMLU
    Q4_K_M
    Context 765K Speed ~29 tok/s

    MoE — speed based on ~23.77 GB active experts plus per-token routing overhead (full 465.28 GB loaded in VRAM)

    Q8_0 Doesn't fit · needs 811.9 GB
  • K2-Horizon 7B 9B Basic Tool useReasoning
    MMLU
    Q4_K_M
    Context 512K Speed ~127 tok/s
    Q8_0
    Context 512K Speed ~79 tok/s
  • Nanbeige4.3 33B 33.12B Basic Tool useReasoning
    MMLU
    Q4_K_M
    Context 512K Speed ~39 tok/s
    Q8_0
    Context 512K Speed ~23 tok/s
  • Nemotron 3 Nano 4B 3.97B Basic ReasoningTool use
    MMLU
    Q4_K_M
    Context 256K Speed ~211 tok/s
    Q8_0
    Context 256K Speed ~161 tok/s
  • Step 3.7 Flash 198B Basic VisionTool useReasoning
    MMLU
    Q4_K_S
    Context 256K Speed ~90 tok/s

    MoE — speed based on ~5.77 GB active experts plus per-token routing overhead (full 103.84 GB loaded in VRAM)

    Q8_0
    Context 256K Speed ~57 tok/s

    MoE — speed based on ~10.84 GB active experts plus per-token routing overhead (full 195.04 GB loaded in VRAM)

  • Ornith 1.5 9B 9.65B Basic Tool useReasoning
    MMLU
    Q4_K_M
    Context 256K Speed ~124 tok/s
    Q8_0
    Context 256K Speed ~78 tok/s
  • Ornith 1.5 35B A3B 35.95B Basic Tool useReasoning
    MMLU
    Q4_K_M
    Context 256K Speed ~171 tok/s

    MoE — speed based on ~1.69 GB active experts plus per-token routing overhead (full 20.22 GB loaded in VRAM)

    Q8_0
    Context 256K Speed ~134 tok/s

    MoE — speed based on ~2.94 GB active experts plus per-token routing overhead (full 35.21 GB loaded in VRAM)

  • Ornith 1.5 397B 403.4B Basic Tool useReasoning
    MMLU
    Q4_K_M
    Context 256K Speed ~66 tok/s

    MoE — speed based on ~9.02 GB active experts plus per-token routing overhead (full 227.53 GB loaded in VRAM)

    Q8_0
    Context 256K Speed ~42 tok/s

    MoE — speed based on ~15.83 GB active experts plus per-token routing overhead (full 399.09 GB loaded in VRAM)

  • Laguna XS 2.1 33.44B Basic Tool useReasoning
    MMLU
    Q4_K_M
    Context 256K Speed ~170 tok/s

    MoE — speed based on ~1.72 GB active experts plus per-token routing overhead (full 19.14 GB loaded in VRAM)

    Q8_0
    Context 256K Speed ~134 tok/s

    MoE — speed based on ~2.97 GB active experts plus per-token routing overhead (full 33.15 GB loaded in VRAM)

  • Nemotron 3.5 Bolt 74B A9B 74.89B Basic VisionTool useReasoning
    MMLU
    Q8_0
    Context 256K Speed ~67 tok/s

    MoE — speed based on ~8.83 GB active experts plus per-token routing overhead (full 73.44 GB loaded in VRAM)

  • Ornith 1.5 122B 125.09B Basic Tool useReasoning
    MMLU
    Q4_K_M
    Context 256K Speed ~92 tok/s

    MoE — speed based on ~5.65 GB active experts plus per-token routing overhead (full 70.64 GB loaded in VRAM)

    Q8_0
    Context 256K Speed ~61 tok/s

    MoE — speed based on ~9.87 GB active experts plus per-token routing overhead (full 123.45 GB loaded in VRAM)

  • Flint 40B A2.5B 39.79B Basic Tool useReasoning
    MMLU
    Q4_K_M
    Context 256K Speed ~179 tok/s

    MoE — speed based on ~1.48 GB active experts plus per-token routing overhead (full 23.61 GB loaded in VRAM)

    Q8_0
    Context 256K Speed ~146 tok/s

    MoE — speed based on ~2.48 GB active experts plus per-token routing overhead (full 39.44 GB loaded in VRAM)

  • MiniCPM-V 4.7 35B A3B 35.21B Basic VisionTool useReasoning
    MMLU
    Q4_K_M
    Context 256K Speed ~171 tok/s

    MoE — speed based on ~1.68 GB active experts plus per-token routing overhead (full 19.71 GB loaded in VRAM)

    Q8_0
    Context 256K Speed ~134 tok/s

    MoE — speed based on ~2.93 GB active experts plus per-token routing overhead (full 34.37 GB loaded in VRAM)

  • MiMo V2.6 Distill Qwen 9B 9.41B Basic VisionTool useReasoning
    MMLU
    Q4_K_M
    Context 256K Speed ~123 tok/s
    Q8_0
    Context 256K Speed ~80 tok/s
  • Ling 3.1 Flash 572.12B Basic Tool useReasoning
    MMLU
    Q4_K_M
    Context 256K Speed ~45 tok/s

    MoE — speed based on ~14.57 GB active experts plus per-token routing overhead (full 333.39 GB loaded in VRAM)

    Q8_0 Doesn't fit · needs 566.6 GB
  • LFM3 1.7B 1.73B Basic Tool useReasoning
    MMLU
    Q4_K_M
    Context 256K Speed ~401 tok/s
    Q8_0
    Context 256K Speed ~295 tok/s
  • Arete 27B 27.36B Basic VisionTool useReasoning
    MMLU
    Q4_K_M
    Context 256K Speed ~46 tok/s
    Q8_0
    Context 256K Speed ~28 tok/s
  • Muse Glimmer 30B 29.78B Basic VisionTool useReasoning
    MMLU
    Q4_K_M
    Context 128K Speed ~47 tok/s
    Q8_0
    Context 128K Speed ~27 tok/s
  • LFM2.5 2.6B 2.7B Basic Tool use
    MMLU
    Q4_K_M
    Context 128K Speed ~314 tok/s
    Q8_0
    Context 128K Speed ~217 tok/s
  • LFM2.5 24B A2B 23.97B Basic Tool useReasoning
    MMLU
    Q4_K_M
    Context 128K Speed ~183 tok/s

    MoE — speed based on ~1.38 GB active experts plus per-token routing overhead (full 13.53 GB loaded in VRAM)

    Q8_0
    Context 128K Speed ~147 tok/s

    MoE — speed based on ~2.43 GB active experts plus per-token routing overhead (full 23.74 GB loaded in VRAM)

  • Mellum2.1 12B A2.5B Thinking 12.15B Basic Tool useReasoning
    MMLU
    Q4_K_M
    Context 128K Speed ~176 tok/s

    MoE — speed based on ~1.55 GB active experts plus per-token routing overhead (full 7.52 GB loaded in VRAM)

    Q8_0
    Context 128K Speed ~146 tok/s

    MoE — speed based on ~2.48 GB active experts plus per-token routing overhead (full 12.04 GB loaded in VRAM)

  • LFM3-VL 2B 1.92B Basic VisionTool useReasoning
    MMLU
    Q4_K_M
    Context 128K Speed ~401 tok/s
    Q8_0
    Context 128K Speed ~295 tok/s
  • Nemotron 3.5 Bolt E4B 8B Basic Tool useReasoning
    MMLU
    Q4_K_M
    Context 64K Speed ~117 tok/s
    Q8_0
    Context 64K Speed ~93 tok/s

Tight fit 2

  • Gemma 2 27B 27.2B Great
    MMLU 75.2
    Q4_K_M
    Context 8K Speed ~48 tok/s

    Only 8K context — under the 16K needed to be practical

    Q8_0
    Context 8K Speed ~28 tok/s

    Only 8K context — under the 16K needed to be practical

  • Gemma 2 9B 9.24B Good
    MMLU 71.3
    Q4_K_M
    Context 8K Speed ~124 tok/s

    Only 8K context — under the 16K needed to be practical

    Q8_0
    Context 8K Speed ~78 tok/s

    Only 8K context — under the 16K needed to be practical

Notify me when new models fit my GPU

An email when a new model fits your M5 Ultra — no spam, unsubscribe anytime.

Doesn't fit 4

  • MiMo V2.5 Pro 1020B Excellent Reasoning
    MMLU 89.4
    Q4_K_M Needs 586.5 GB — 74.5 GB over
    Q8_0 Needs 1013.0 GB — 501.0 GB over
  • DeepSeek V4 Pro 1600B Excellent Tool useReasoning
    MMLU 87.5
    Q4_K_M Needs 900.1 GB — 388.1 GB over
    Q8_0 Needs 1670.1 GB — 1158.1 GB over
  • GLM 5.3 753.33B Excellent Tool useReasoning
    MMLU 86.8
    Q8_0 Needs 746.4 GB — 234.4 GB over

Doesn't fit 4

  • MiMo V2.5 Pro 1020B Excellent Reasoning
    MMLU 89.4
    Q4_K_M Needs 586.5 GB — 74.5 GB over
    Q8_0 Needs 1013.0 GB — 501.0 GB over
  • DeepSeek V4 Pro 1600B Excellent Tool useReasoning
    MMLU 87.5
    Q4_K_M Needs 900.1 GB — 388.1 GB over
    Q8_0 Needs 1670.1 GB — 1158.1 GB over
  • GLM 5.3 753.33B Excellent Tool useReasoning
    MMLU 86.8
    Q8_0 Needs 746.4 GB — 234.4 GB over
  • Mistral Large 4 1T A52B 1049.43B Basic VisionTool useReasoning
    MMLU
    Q4_K_M Needs 591.3 GB — 79.3 GB over
    Q8_0 Needs 1038.6 GB — 526.6 GB over