Mellum2.1 12B A2.5B Thinking VRAM requirements: GPUs that run it
The smallest GPU memory that runs Mellum2.1 12B A2.5B Thinking well is 8 GB (Q4_K_M, e.g. GeForce RTX 5060 Ti). 120 of 132 GPUs run it well. It's a mixture-of-experts model, so all weights must fit in memory but speed depends only on the active experts.
SWE-bench Verified 47
VRAM needed per quantization
| Quantization | Weights | 16K context | 32K context | Full 128K |
|---|---|---|---|---|
| Q4_K_M | 7.5 GB | 7.7 GB | 7.9 GB | 9.2 GB |
| Q8_0 | 12.0 GB | 12.2 GB | 12.5 GB | 13.7 GB |
Weights plus fp16 KV cache. Leave ~1 GB headroom for the inference engine.
Run it locally
Each command downloads the GGUF from Hugging Face, then Ollama starts a chat and llama.cpp starts an OpenAI-compatible server with a web UI on port 8080.
ollama run hf.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking-GGUF:Q4_K_M llama-server -hf JetBrains/Mellum2.1-12B-A2.5B-Thinking-GGUF:Q4_K_M ollama run hf.co/JetBrains/Mellum2.1-12B-A2.5B-Thinking-GGUF:Q8_0 llama-server -hf JetBrains/Mellum2.1-12B-A2.5B-Thinking-GGUF:Q8_0 Cheapest GPUs that can run Mellum2.1 12B A2.5B Thinking
Lowest launch MSRP among GPUs that run its smallest quantization (Q4_K_M) well, with at least 16K context.
As an Amazon Associate, I earn from qualifying purchases.
GPUs that can run Mellum2.1 12B A2.5B Thinking
Best quantization that fits on each GPU, with the context and speed it gets. For chips with several memory sizes, the smallest that works is shown.
NVIDIA
- GTX 1070 8 GB Q4_K_M 36K context · ~77 tok/s
- GTX 1070 Ti 8 GB Q4_K_M 36K context · ~77 tok/s
- GTX 1080 8 GB Q4_K_M 36K context · ~92 tok/s
- RTX 2060 SUPER 8 GB Q4_K_M 36K context · ~118 tok/s
- RTX 2070 8 GB Q4_K_M 36K context · ~118 tok/s
- RTX 2070 SUPER 8 GB Q4_K_M 36K context · ~118 tok/s
- RTX 2080 8 GB Q4_K_M 36K context · ~118 tok/s
- RTX 2080 SUPER 8 GB Q4_K_M 36K context · ~127 tok/s
- RTX 3050 8 GB Q4_K_M 36K context · ~69 tok/s
- RTX 3060 Ti 8 GB Q4_K_M 36K context · ~118 tok/s
- RTX 3060 Ti GDDR6X 8 GB Q4_K_M 36K context · ~145 tok/s
- RTX 3070 8 GB Q4_K_M 36K context · ~118 tok/s
- RTX 3070 Ti 8 GB Q4_K_M 36K context · ~145 tok/s
- RTX 4060 8 GB Q4_K_M 36K context · ~81 tok/s
- RTX 4060 Ti 8 GB Q4_K_M 36K context · ~85 tok/s
- RTX 5050 8 GB Q4_K_M 36K context · ~92 tok/s
- RTX 5060 8 GB Q4_K_M 36K context · ~118 tok/s
- RTX 5060 Ti 8 GB Q4_K_M 36K context · ~118 tok/s
- T1000 8GB 8 GB Q4_K_M 36K context · ~52 tok/s
- RTX 3080 10 GB Q4_K_M 128K context · ~167 tok/s
- GTX 1080 Ti 11 GB Q4_K_M 128K context · ~125 tok/s
- RTX 2080 Ti 11 GB Q4_K_M 128K context · ~147 tok/s
- RTX 2060 12GB 12 GB Q4_K_M 128K context · ~96 tok/s
- RTX 3060 12 GB Q4_K_M 128K context · ~101 tok/s
- RTX 3080 12GB 12 GB Q4_K_M 128K context · ~184 tok/s
- RTX 3080 Ti 12 GB Q4_K_M 128K context · ~184 tok/s
- RTX 4070 12 GB Q4_K_M 128K context · ~128 tok/s
- RTX 4070 SUPER 12 GB Q4_K_M 128K context · ~128 tok/s
- RTX 4070 Ti 12 GB Q4_K_M 128K context · ~128 tok/s
- RTX 5070 12 GB Q4_K_M 128K context · ~155 tok/s
- RTX A2000 12GB 12 GB Q4_K_M 128K context · ~85 tok/s
- RTX 4060 Ti 16GB 16 GB Q8_0 128K context · ~66 tok/s
- RTX 4070 Ti SUPER 16 GB Q8_0 128K context · ~126 tok/s
- RTX 4080 16 GB Q8_0 128K context · ~131 tok/s
- RTX 4080 SUPER 16 GB Q8_0 128K context · ~134 tok/s
- RTX 5060 Ti 16GB 16 GB Q8_0 128K context · ~94 tok/s
- RTX 5070 Ti 16 GB Q8_0 128K context · ~152 tok/s
- RTX 5080 16 GB Q8_0 128K context · ~158 tok/s
- RTX 2000 Ada 16 GB Q8_0 128K context · ~53 tok/s
- RTX A4000 16 GB Q8_0 128K context · ~94 tok/s
- RTX PRO 2000 Blackwell 16 GB Q8_0 128K context · ~66 tok/s
- Tesla T4 16 GB Q8_0 128K context · ~72 tok/s
- Tesla V100 16GB 16 GB Q8_0 128K context · ~152 tok/s
- RTX 4000 Ada 20 GB Q8_0 128K context · ~79 tok/s
- RTX 3090 24 GB Q8_0 128K context · ~156 tok/s
- RTX 3090 Ti 24 GB Q8_0 128K context · ~163 tok/s
- RTX 4090 24 GB Q8_0 128K context · ~163 tok/s
- RTX 4500 Ada 24 GB Q8_0 128K context · ~91 tok/s
- RTX A5000 24 GB Q8_0 128K context · ~138 tok/s
- RTX PRO 4000 Blackwell 24 GB Q8_0 128K context · ~126 tok/s
- RTX Spark 24 GB Q8_0 128K context · ~68 tok/s
- Tesla P40 24 GB Q8_0 128K context · ~76 tok/s
- RTX 5090 32 GB Q8_0 128K context · ~220 tok/s
- RTX 5000 Ada 32 GB Q8_0 128K context · ~113 tok/s
- RTX PRO 4500 Blackwell 32 GB Q8_0 128K context · ~152 tok/s
- Tesla V100 32GB 32 GB Q8_0 128K context · ~152 tok/s
- L40S 48 GB Q8_0 128K context · ~148 tok/s
- RTX 6000 Ada 48 GB Q8_0 128K context · ~158 tok/s
- RTX A6000 48 GB Q8_0 128K context · ~138 tok/s
- RTX PRO 5000 Blackwell 48 GB Q8_0 128K context · ~191 tok/s
- RTX PRO 5000 Blackwell 72GB 72 GB Q8_0 128K context · ~191 tok/s
- RTX PRO 5500 Blackwell 84 GB Q8_0 128K context · ~195 tok/s
- RTX PRO 6000 Blackwell 96 GB Q8_0 128K context · ~220 tok/s
- DGX Spark (GB 10) 128 GB Q8_0 128K context · ~63 tok/s
AMD
- Radeon RX 5700 8 GB Q4_K_M 36K context · ~118 tok/s
- Radeon RX 5700 XT 8 GB Q4_K_M 36K context · ~118 tok/s
- Radeon RX 6600 8 GB Q4_K_M 36K context · ~69 tok/s
- Radeon RX 6600 XT 8 GB Q4_K_M 36K context · ~77 tok/s
- Radeon RX 6650 XT 8 GB Q4_K_M 36K context · ~83 tok/s
- Radeon RX 7600 8 GB Q4_K_M 36K context · ~85 tok/s
- Radeon RX 9060 8 GB Q4_K_M 36K context · ~85 tok/s
- Radeon RX 9060 XT 8 GB Q4_K_M 36K context · ~92 tok/s
- Radeon RX 6700 XT 12 GB Q4_K_M 128K context · ~106 tok/s
- Radeon RX 6750 XT 12 GB Q4_K_M 128K context · ~115 tok/s
- Radeon RX 7700 XT 12 GB Q4_K_M 128K context · ~115 tok/s
- Radeon RX 9070 GRE 12 GB Q4_K_M 128K context · ~115 tok/s
- AMD BC-250 16 GB Q8_0 128K context · ~94 tok/s
- Radeon RX 6800 16 GB Q8_0 128K context · ~104 tok/s
- Radeon RX 6800 XT 16 GB Q8_0 128K context · ~104 tok/s
- Radeon RX 6900 XT 16 GB Q8_0 128K context · ~104 tok/s
- Radeon RX 6950 XT 16 GB Q8_0 128K context · ~113 tok/s
- Radeon RX 7600 XT 16 GB Q8_0 128K context · ~66 tok/s
- Radeon RX 7800 XT 16 GB Q8_0 128K context · ~119 tok/s
- Radeon RX 7900 GRE 16 GB Q8_0 128K context · ~113 tok/s
- Radeon RX 9060 XT 16GB 16 GB Q8_0 128K context · ~72 tok/s
- Radeon RX 9070 16 GB Q8_0 128K context · ~122 tok/s
- Radeon RX 9070 XT 16 GB Q8_0 128K context · ~122 tok/s
- Radeon RX 7900 XT 20 GB Q8_0 128K context · ~141 tok/s
- Radeon RX 7900 XTX 24 GB Q8_0 128K context · ~158 tok/s
- Radeon AI PRO R9700 32 GB Q8_0 128K context · ~122 tok/s
- Ryzen AI Max (Strix Halo) 32 GB Q8_0 128K context · ~59 tok/s
- Ryzen AI Max 400 (Gorgon Halo) 64 GB Q8_0 128K context · ~63 tok/s
Intel
- Arc A580 8 GB Q4_K_M 36K context · ~130 tok/s
- Arc A750 8 GB Q4_K_M 36K context · ~130 tok/s
- Arc A770 8GB 8 GB Q4_K_M 36K context · ~130 tok/s
- Arc B570 10 GB Q4_K_M 128K context · ~105 tok/s
- Arc B580 12 GB Q4_K_M 128K context · ~120 tok/s
- Arc A770 16GB 16 GB Q8_0 128K context · ~111 tok/s
- Arc Pro B65 32 GB Q8_0 128K context · ~117 tok/s
- Arc Pro B70 32 GB Q8_0 128K context · ~117 tok/s
Apple
- M1 8 GB Q4_K_M 36K context · ~24 tok/s
- M2 8 GB Q4_K_M 36K context · ~34 tok/s
- M3 8 GB Q4_K_M 36K context · ~34 tok/s
- M1 Pro 16 GB Q8_0 128K context · ~48 tok/s
- M2 Pro 16 GB Q8_0 128K context · ~48 tok/s
- M4 16 GB Q8_0 128K context · ~30 tok/s
- M5 16 GB Q8_0 128K context · ~38 tok/s
- M6 16 GB Q8_0 128K context · ~38 tok/s
- M3 Pro 18 GB Q8_0 128K context · ~37 tok/s
- M4 Pro 24 GB Q8_0 128K context · ~63 tok/s
- M5 Pro 24 GB Q8_0 128K context · ~69 tok/s
- M1 Max 32 GB Q8_0 128K context · ~86 tok/s
- M2 Max 32 GB Q8_0 128K context · ~86 tok/s
- M3 Max 36 GB Q8_0 128K context · ~68 tok/s
- M4 Max 36 GB Q8_0 128K context · ~87 tok/s
- M5 Max 36 GB Q8_0 128K context · ~96 tok/s
- M1 Ultra 64 GB Q8_0 128K context · ~111 tok/s
- M2 Ultra 64 GB Q8_0 128K context · ~111 tok/s
- M3 Ultra 96 GB Q8_0 128K context · ~112 tok/s
- M5 Ultra 96 GB Q8_0 128K context · ~146 tok/s