Our catalog lists 7 GPUs with 12 GB VRAM (RTX 5070 12GB, RTX 4070 Ti 12GB, RTX 4070 12GB, RTX 4070 Super 12GB, Arc B580 12GB and 2 more). The ranking uses the RTX 5070 12GB — the highest-bandwidth 12 GB card at 672 GB/s — and shows the 18 of 61 catalog models that fully fit it at Q4 / 8K context (weights + KV cache + 1 GB overhead within 95% of VRAM), fastest first. Methodology.
| # | Model | Params (B) | Q4 tok/s | VRAM needed (GB) | Fit check |
|---|---|---|---|---|---|
| 1 | Qwen3-0.6B | 0.8 | 534.5 est. | 1.6 | fit check |
| 2 | gemma-3-1b-it | 1.0 | 427.6 est. | 1.75 | fit check |
| 3 | Llama-3.2-1B-Instruct | 1.2 | 356.4 est. | 1.9 | fit check |
| 4 | Qwen3-1.7B | 2.0 | 213.8 est. | 2.5 | fit check |
| 5 | Llama-3.2-3B-Instruct | 3.2 | 133.6 est. | 3.4 | fit check |
| 6 | Mistral-7B-Instruct-v0.3 | 7.2 | 120.5 measured | 6.4 | fit check |
| 7 | Phi-4-mini-instruct | 3.8 | 112.5 est. | 3.85 | fit check |
| 8 | Phi-4-mini-reasoning | 3.8 | 112.5 est. | 3.85 | fit check |
| 9 | gemma-3-4b-it | 4.3 | 109.7 est. | 4.15 | fit check |
| 10 | Qwen3-4B-Instruct-2507 | 4.0 | 106.9 est. | 4.0 | fit check |
| 11 | Qwen3-4B-Thinking-2507 | 4.0 | 106.9 est. | 4.0 | fit check |
| 12 | Llama-3.1-8B-Instruct | 8.0 | 100.4 measured | 7.0 | fit check |
| 13 | Qwen2.5-7B-Instruct | 7.6 | 56.3 est. | 6.7 | fit check |
| 14 | Qwen2.5-Coder-7B-Instruct | 7.6 | 56.3 est. | 6.7 | fit check |
| 15 | DeepSeek-R1-Distill-Qwen-7B | 7.6 | 56.3 est. | 6.7 | fit check |
| 16 | Qwen3-8B | 8.2 | 52.2 est. | 7.15 | fit check |
| 17 | DeepSeek-R1-0528-Qwen3-8B | 8.2 | 52.2 est. | 7.15 | fit check |
| 18 | gemma-3-12b-it | 12.2 | 35.6 est. | 10.11 | fit check |