44 models from the tokcalc catalog. Each row shows the VRAM the weights need at the recommended quantization, whether it fits on a single GPU, and the cheapest configuration that serves it — all computed by the same formulas the calculator uses.
| Model | Params | VRAM (weights) | Quant | Fits on 1 GPU | Cheapest config | Cost / 1M tok |
|---|---|---|---|---|---|---|
| Kimi K3 (MoE, VLM)MoE | 2779.93B | 1529.0 GB | AWQ 4-bit | — | 8× AMD MI300X 192GB | $0.1606 |
| GLM-5.3 (MoE)MoE | 753.33B | 414.3 GB | AWQ 4-bit | — | 8× A100 80GB SXM4 | $0.0375 |
| DeepSeek R1 671B (MoE)MoE | 671B | 369.1 GB | AWQ 4-bit | — | 8× A100 80GB SXM4 | $0.0727 |
| DeepSeek V3 671B (MoE)MoE | 671B | 369.1 GB | AWQ 4-bit | — | 8× A100 80GB SXM4 | $0.0727 |
| Llama 3.1 405B | 405B | 222.8 GB | AWQ 4-bit | ✓ | 8× A100 40GB SXM4 | $0.7697 |
| Llama 4 Maverick 17B-128E (MoE, VLM)MoE | 400B | 220.0 GB | AWQ 4-bit | ✓ | 8× RTX 5090 32GB | $0.0155 |
| DeepSeek Coder V2 236B (MoE)MoE | 236B | 129.8 GB | AWQ 4-bit | ✓ | 8× RTX 5090 32GB | $0.0192 |
| Mistral Large 3 (256K)MoE | 235B | 129.3 GB | AWQ 4-bit | ✓ | 8× RTX 5090 32GB | $0.0648 |
| Qwen 3 235B-A22B (MoE)MoE | 235B | 129.3 GB | AWQ 4-bit | ✓ | 8× RTX 5090 32GB | $0.0201 |
| Mixtral 8x22B (MoE)MoE | 141B | 77.6 GB | AWQ 4-bit | ✓ | 8× RTX 4080 16GB | $0.0534 |
| GPT-OSS 120B (MoE)MoE | 116.83B | 64.3 GB | AWQ 4-bit | ✓ | 8× RTX 3080 10GB | $0.004393 |
| Llama 4 Scout 17B-16E (MoE, VLM)MoE | 109B | 60.0 GB | AWQ 4-bit | ✓ | 8× RTX 3080 10GB | $0.0146 |
| Qwen 2.5 72B | 72.70B | 72.7 GB | FP8 (E4M3) | ✓ | 8× RTX 4080 16GB | $0.1062 |
| Qwen2 72B | 72.70B | 72.7 GB | FP8 (E4M3) | ✓ | 8× RTX 4080 16GB | $0.1062 |
| Qwen 2.5-VL 72B (VLM) | 72B | 72.0 GB | FP8 (E4M3) | ✓ | 8× RTX 4080 16GB | $0.1052 |
| Llama 3.3 70B | 70.60B | 70.6 GB | FP8 (E4M3) | ✓ | 8× RTX 4080 16GB | $0.1031 |
| Llama 3 70B | 70.60B | 70.6 GB | FP8 (E4M3) | ✓ | 8× RTX 4080 16GB | $0.1031 |
| Mixtral 8x7B (MoE)MoE | 46.70B | 46.7 GB | FP8 (E4M3) | ✓ | 8× RTX 3080 10GB | $0.0119 |
| Qwen 3 32B | 32B | 32.0 GB | FP8 (E4M3) | ✓ | 8× RTX 3080 10GB | $0.0294 |
| Qwen 3 30B-A3B (MoE)MoE | 30B | 30.0 GB | FP8 (E4M3) | ✓ | 4× RTX 3080 10GB | $0.005513 |
| Qwen3.8 27B (VLM) | 27.78B | 27.8 GB | FP8 (E4M3) | ✓ | 4× RTX 3080 10GB | $0.0511 |
| Gemma 2 27B | 27B | 27.0 GB | FP8 (E4M3) | ✓ | 8× RTX 3080 10GB | $0.0248 |
| Codestral 25.08 | 22B | 22.0 GB | FP8 (E4M3) | ✓ | 4× RTX 3080 10GB | $0.0404 |
| GPT-OSS 20B (MoE)MoE | 20.91B | 20.9 GB | FP8 (E4M3) | ✓ | 4× RTX 3080 10GB | $0.006615 |
| GPT-NeoX 20B | 20B | 20.0 GB | FP8 (E4M3) | ✓ | 8× RTX 4080 16GB | $0.0292 |
| Phi-4 14B | 14B | 28.0 GB | FP16 | ✓ | 4× RTX 3080 10GB | $0.0772 |
| Qwen 2.5 14B | 14B | 28.0 GB | FP16 | ✓ | 8× RTX 3080 10GB | $0.0386 |
| Qwen 3 14B | 14B | 28.0 GB | FP16 | ✓ | 4× RTX 3080 10GB | $0.0772 |
| OLMo 2 13B | 13B | 26.0 GB | FP16 | ✓ | 4× RTX 3080 10GB | $0.0717 |
| Pixtral 12B (VLM) | 12B | 24.0 GB | FP16 | ✓ | 4× RTX 3080 10GB | $0.0662 |
| Falcon 3 10B Instruct | 10B | 20.0 GB | FP16 | ✓ | 4× RTX 3080 10GB | $0.0551 |
| Gemma 2 9B | 9B | 18.0 GB | FP16 | ✓ | 8× RTX 3080 10GB | $0.0248 |
| Llama 3 8B | 8.03B | 16.1 GB | FP16 | ✓ | 4× RTX 3080 10GB | $0.0443 |
| Qwen 3 8B | 8B | 16.0 GB | FP16 | ✓ | 4× RTX 3080 10GB | $0.0441 |
| Qwen2 7B | 7.62B | 15.2 GB | FP16 | ✓ | 2× RTX 3080 10GB | $0.0840 |
| Mistral 7B v0.3 | 7.24B | 14.5 GB | FP16 | ✓ | 4× RTX 3080 10GB | $0.0399 |
| Qwen 2.5-VL 7B (VLM) | 7B | 14.0 GB | FP16 | ✓ | 2× RTX 3080 10GB | $0.0772 |
| Llama 2 7B | 6.74B | 13.5 GB | FP16 | ✓ | 8× RTX 3080 10GB | $0.0186 |
| Qwen 3 4B | 4B | 8.0 GB | FP16 | ✓ | 2× RTX 3080 10GB | $0.0441 |
| Phi-3 Mini 3.8B | 3.80B | 7.6 GB | FP16 | ✓ | 4× RTX 3080 10GB | $0.0209 |
| SmolLM2 1.7B Instruct | 1.70B | 3.4 GB | FP16 | ✓ | 2× RTX 3080 10GB | $0.0187 |
| BGE-M3 (embed, 1024-dim) | 0.57B | 1.1 GB | FP16 | ✓ | 1× RTX 3080 10GB | $0.0106 |
| Multilingual E5 Large (1024-dim) | 0.56B | 1.1 GB | FP16 | ✓ | 1× RTX 3080 10GB | $0.0105 |
| GTE Large (1024-dim) | 0.44B | 0.9 GB | FP16 | ✓ | 1× RTX 3080 10GB | $0.008247 |
Looking for a GPU-first view? See all GPUs and what they can serve.