Best LLMs to run on GB200 NVL36 (186 GB)
datacenter 186 GB HBM3e · 8000 GB/s · 2500 TFLOPS BF16 · updated 2026-10-23
Headline answer
Of 141 models in our catalog, 115 fit on a single GB200 NVL36 at FP8, 122 at 4-bit (AWQ) and 106 at BF16 (8K context, batch 1). Largest that fits: Qwen3-Next 80B-A3B at BF16; Mixtral 8x22B at FP8; MiMo V2-Flash at 4-bit (AWQ). Rental: $5.50/hr catalog estimate.
Change context, concurrency or framework in the calculator → GB200 NVL36 full specs
Every model on GB200 NVL36
Fit band and share of VRAM per precision; N× GPUs = minimum count inside one node (tensor parallel); — = does not fit within one 8-GPU node. Speed and $/1M tokens for the highest-quality precision that fits on one GPU (BF16, else FP8, else 4-bit), single stream.
| Model | Params | BF16 | FP8 | 4-bit (AWQ) | tok/s | $ / 1M tok |
|---|---|---|---|---|---|---|
| MiMo V2-Flash | 309.8B (15.0B active) | 4× GPUs | 2× GPUs | tight 90% | 218 4-bit | $7.00 |
| DeepSeek V4-Flash | 284.0B (13.0B active) | 4× GPUs | 2× GPUs | tight 83% | 229 4-bit | $6.67 |
| Llama-3.3-Nemotron Ultra 253B | 253.0B | 4× GPUs | 2× GPUs | good 79% | 34 4-bit | $44.93 |
| DeepSeek Coder V2 | 236.0B (21.0B active) | 4× GPUs | 2× GPUs | optimal 69% | 192 4-bit | $7.94 |
| Qwen3 235B-A22B | 235.0B (22.0B active) | 4× GPUs | 2× GPUs | optimal 70% | 188 4-bit | $8.13 |
| MiniMax M2.5 | 230.0B (10.0B active) | 4× GPUs | 2× GPUs | optimal 69% | 245 4-bit | $6.24 |
| MiniMax M2.7 | 230.0B (10.0B active) | 4× GPUs | 2× GPUs | optimal 69% | 245 4-bit | $6.24 |
| Mixtral 8x22B | 141.0B (39.0B active) | 2× GPUs | tight 81% | optimal 43% | 90 FP8 | $17.05 |
| DBRX 132B | 132.0B (36.0B active) | 2× GPUs | good 75% | optimal 41% | 95 FP8 | $16.08 |
| Devstral 2 123B | 123.0B | 2× GPUs | good 71% | optimal 41% | 35 FP8 | $43.78 |
| Mistral Large 2 123B | 123.0B | 2× GPUs | good 71% | optimal 39% | 35 FP8 | $43.78 |
| Nemotron-3 Super 120B-A12B | 120.0B (12.0B active) | 2× GPUs | optimal 68% | optimal 37% | 182 FP8 | $8.41 |
| Mistral Small 4 119B | 119.0B (11.0B active) | 2× GPUs | optimal 68% | optimal 36% | 189 FP8 | $8.09 |
| GPT OSS 120B | 117.0B (5.1B active) | 2× GPUs | optimal 67% | optimal 36% | 245 FP8 | $6.23 |
| Llama 4 Scout (109B) | 109.0B (17.0B active) | 2× GPUs | optimal 63% | optimal 35% | 152 FP8 | $10.04 |
| Command R+ 104B | 104.0B | 2× GPUs | optimal 60% | optimal 38% | 40 FP8 | $37.72 |
| Qwen3-Next 80B-A3B | 80.0B (3.0B active) | tight 89% | optimal 46% | optimal 26% | 234 BF16 | $6.52 |
| Qwen 2 72B | 72.0B | tight 81% | optimal 43% | optimal 26% | 30 BF16 | $50.42 |
| Qwen 2.5 72B | 72.0B | tight 81% | optimal 43% | optimal 26% | 30 BF16 | $50.42 |
| CodeLlama 70B | 70.0B | good 79% | optimal 42% | optimal 24% | 31 BF16 | $49.28 |
| DeepSeek R1 Distill 70B | 70.0B | good 79% | optimal 42% | optimal 25% | 31 BF16 | $49.28 |
| LLaMA 2 70B 4K ctx | 70.0B | good 78% | optimal 41% | optimal 23% | 31 BF16 | $49.28 |
| LLaMA 3.1 70B | 70.0B | good 79% | optimal 42% | optimal 25% | 31 BF16 | $49.28 |
| LLaMA 3.3 70B | 70.0B | good 79% | optimal 42% | optimal 25% | 31 BF16 | $49.28 |
| Llama-3.1-Nemotron-70B | 70.0B | good 79% | optimal 42% | optimal 25% | 31 BF16 | $49.28 |
| Nemotron-H 56B Base | 56.0B | optimal 64% | optimal 35% | optimal 22% | 38 BF16 | $40.31 |
| Llama-3.1-Nemotron-51B | 51.0B | optimal 58% | optimal 31% | optimal 19% | 41 BF16 | $37.08 |
| Llama-3.3-Nemotron Super 49B | 49.0B | optimal 56% | optimal 30% | optimal 18% | 43 BF16 | $35.86 |
| Kimi Linear 48B-A3B | 48.0B (3.0B active) | optimal 54% | optimal 29% | optimal 16% | 235 BF16 | $6.50 |
| Mixtral 8x7B | 46.7B (12.9B active) | optimal 53% | optimal 28% | optimal 16% | 119 BF16 | $12.83 |
| Command R 35B | 35.0B | optimal 46% | optimal 27% | optimal 21% | 56 BF16 | $27.18 |
| Qwen3.6 35B-A3B | 35.0B (3.0B active) | optimal 40% | optimal 22% | optimal 13% | 235 BF16 | $6.51 |
| CodeLlama 34B | 34.0B | optimal 40% | optimal 22% | optimal 13% | 58 BF16 | $26.30 |
| Yi-1.5 34B 4K ctx | 34.0B | optimal 39% | optimal 21% | optimal 13% | 58 BF16 | $26.30 |
| DeepSeek R1 Distill 32B | 32.0B | optimal 38% | optimal 21% | optimal 14% | 61 BF16 | $25.00 |
| Nemotron-Terminal-32B | 32.0B | optimal 38% | optimal 21% | optimal 14% | 61 BF16 | $25.00 |
| OLMo-2 32B 4K ctx | 32.0B | optimal 37% | optimal 20% | optimal 12% | 61 BF16 | $25.00 |
| Qwen 2.5 32B | 32.0B | optimal 38% | optimal 21% | optimal 14% | 61 BF16 | $25.00 |
| Qwen 2.5 Coder 32B | 32.0B | optimal 38% | optimal 21% | optimal 14% | 61 BF16 | $25.00 |
| Qwen3 32B | 32.0B | optimal 38% | optimal 21% | optimal 14% | 61 BF16 | $25.00 |
| Nemotron-3 Nano 30B-A3B | 31.6B (3.2B active) | optimal 36% | optimal 20% | optimal 12% | 231 BF16 | $6.62 |
| Muse Glimmer 30B | 30.0B | optimal 35% | optimal 19% | optimal 12% | 64 BF16 | $23.69 |
| Qwen3 30B-A3B | 30.0B (3.0B active) | optimal 35% | optimal 19% | optimal 12% | 234 BF16 | $6.52 |
| Gemma 2 27B | 27.0B | optimal 33% | optimal 19% | optimal 13% | 70 BF16 | $21.86 |
| Gemma 3 27B | 27.0B | optimal 34% | optimal 19% | optimal 14% | 70 BF16 | $21.89 |
| Qwen3.6 27B | 27.0B | optimal 32% | optimal 18% | optimal 13% | 70 BF16 | $21.83 |
| Devstral 24B | 24.0B | optimal 29% | optimal 16% | optimal 11% | 77 BF16 | $19.89 |
| Devstral Small 2 24B | 24.0B | optimal 29% | optimal 16% | optimal 11% | 77 BF16 | $19.89 |
| Magistral Small 24B | 24.0B | optimal 29% | optimal 16% | optimal 11% | 77 BF16 | $19.89 |
| Mistral Small 3.1 24B | 24.0B | optimal 29% | optimal 16% | optimal 11% | 77 BF16 | $19.89 |
| Codestral 22B | 22.0B | optimal 27% | optimal 15% | optimal 10% | 82 BF16 | $18.65 |
| GPT OSS 20B | 21.0B (3.6B active) | optimal 25% | optimal 14% | optimal 9% | 222 BF16 | $6.89 |
| InternLM 2.5 20B | 20.0B | optimal 25% | optimal 14% | optimal 10% | 88 BF16 | $17.36 |
| DeepSeek Coder V2 Lite | 16.0B (2.4B active) | optimal 20% | optimal 11% | optimal 7% | 250 BF16 | $6.12 |
| StarCoder2 15B | 16.0B | optimal 20% | optimal 11% | optimal 8% | 104 BF16 | $14.76 |
| DeepSeek R1 Distill 14B | 14.8B | optimal 19% | optimal 11% | optimal 8% | 109 BF16 | $14.03 |
| Qwen 2.5 14B | 14.8B | optimal 19% | optimal 11% | optimal 8% | 109 BF16 | $14.03 |
| Qwen3 14B | 14.8B | optimal 19% | optimal 11% | optimal 8% | 109 BF16 | $14.03 |
| Phi 3 Medium | 14.0B | optimal 18% | optimal 11% | optimal 7% | 113 BF16 | $13.54 |
| Phi-4 14B | 14.0B | optimal 18% | optimal 11% | optimal 8% | 113 BF16 | $13.54 |
| Qwen 2.5 Coder 14B | 14.0B | optimal 18% | optimal 11% | optimal 8% | 113 BF16 | $13.54 |
| OLMo-2 13B 4K ctx | 13.7B | optimal 19% | optimal 12% | optimal 9% | 113 BF16 | $13.52 |
| LLaMA 2 13B 4K ctx | 13.0B | optimal 18% | optimal 11% | optimal 8% | 117 BF16 | $13.06 |
| Gemma 3 12B | 12.0B | optimal 17% | optimal 10% | optimal 9% | 124 BF16 | $12.32 |
| Mistral Nemo 12B | 12.0B | optimal 16% | optimal 10% | optimal 7% | 125 BF16 | $12.26 |
| Nemotron-Nano-12B-v2 | 12.0B | optimal 16% | optimal 10% | optimal 8% | 124 BF16 | $12.28 |
| StableLM 2 12B 4K ctx | 12.0B | optimal 16% | optimal 9% | optimal 7% | 124 BF16 | $12.27 |
| Falcon3 10B | 10.0B | optimal 14% | optimal 8% | optimal 6% | 139 BF16 | $10.99 |
| GLM-4 9B | 9.0B | optimal 12% | optimal 7% | optimal 6% | 148 BF16 | $10.32 |
| Gemma 2 9B | 9.0B | optimal 13% | optimal 9% | optimal 8% | 147 BF16 | $10.40 |
| Nemotron-Nano-9B-v2 | 9.0B | optimal 13% | optimal 8% | optimal 7% | 147 BF16 | $10.37 |
| Yi-1.5 9B 4K ctx | 9.0B | optimal 12% | optimal 7% | optimal 5% | 148 BF16 | $10.34 |
| InternLM3 8B | 8.8B | optimal 12% | optimal 7% | optimal 6% | 150 BF16 | $10.20 |
| Qwen3-VL 8B | 8.8B | optimal 12% | optimal 8% | optimal 6% | 150 BF16 | $10.20 |
| MiniCPM-V 4.5 | 8.7B | optimal 12% | optimal 8% | optimal 6% | 150 BF16 | $10.16 |
| Ministral-3 8B | 8.5B | optimal 12% | optimal 8% | optimal 6% | 152 BF16 | $10.03 |
| Qwen2.5-VL 7B | 8.3B | optimal 11% | optimal 7% | optimal 6% | 155 BF16 | $9.88 |
| DeepSeek R1 Distill 8B | 8.0B | optimal 11% | optimal 7% | optimal 6% | 157 BF16 | $9.71 |
| Granite 3.3 8B | 8.0B | optimal 12% | optimal 7% | optimal 6% | 157 BF16 | $9.72 |
| LLaMA 3.1 8B | 8.0B | optimal 11% | optimal 7% | optimal 6% | 157 BF16 | $9.71 |
| Ministral 8B | 8.0B | optimal 11% | optimal 7% | optimal 6% | 157 BF16 | $9.71 |
| Qwen3 8B | 8.0B | optimal 11% | optimal 7% | optimal 6% | 157 BF16 | $9.71 |
| Qwen 2.5 7B | 7.6B | optimal 11% | optimal 7% | optimal 5% | 162 BF16 | $9.45 |
| Falcon3 7B | 7.5B | optimal 11% | optimal 7% | optimal 5% | 163 BF16 | $9.36 |
| Phi-3 Small 7B | 7.4B | optimal 11% | optimal 7% | optimal 6% | 164 BF16 | $9.32 |
| InternLM 2.5 7B | 7.0B | optimal 10% | optimal 7% | optimal 5% | 168 BF16 | $9.07 |
| LLaMA 2 7B 4K ctx | 7.0B | optimal 11% | optimal 7% | optimal 6% | 167 BF16 | $9.16 |
| Mistral 7B | 7.0B | optimal 10% | optimal 7% | optimal 5% | 168 BF16 | $9.07 |
| OLMo-2 7B 4K ctx | 7.0B | optimal 11% | optimal 7% | optimal 6% | 167 BF16 | $9.16 |
| Qwen 2 7B | 7.0B | optimal 10% | optimal 6% | optimal 5% | 169 BF16 | $9.05 |
| Qwen 2.5 Coder 7B | 7.0B | optimal 10% | optimal 6% | optimal 5% | 169 BF16 | $9.05 |
| StarCoder2 7B | 7.0B | optimal 10% | optimal 6% | optimal 5% | 169 BF16 | $9.06 |
| Yi-1.5 6B 4K ctx | 6.0B | optimal 9% | optimal 6% | optimal 4% | 182 BF16 | $8.42 |
| Qwen3-VL 4B | 4.4B | optimal 8% | optimal 5% | optimal 5% | 205 BF16 | $7.45 |
| Gemma 3 4B | 4.3B | optimal 7% | optimal 5% | optimal 5% | 208 BF16 | $7.36 |
| Nemotron-Mini-4B 4K ctx | 4.0B | optimal 7% | optimal 5% | optimal 5% | 213 BF16 | $7.16 |
| Qwen3 4B | 4.0B | optimal 7% | optimal 5% | optimal 5% | 213 BF16 | $7.17 |
| Ministral 3B | 3.8B | optimal 7% | optimal 5% | optimal 4% | 218 BF16 | $7.02 |
| Phi 3.5 Mini | 3.8B | optimal 8% | optimal 6% | optimal 5% | 215 BF16 | $7.10 |
| Phi-4 Mini | 3.8B | optimal 7% | optimal 5% | optimal 5% | 217 BF16 | $7.03 |
| Qwen2.5-VL 3B | 3.8B | optimal 6% | optimal 4% | optimal 4% | 219 BF16 | $6.98 |
| Ministral-3 3B | 3.4B | optimal 6% | optimal 4% | optimal 4% | 226 BF16 | $6.78 |
| LLaMA 3.2 3B | 3.2B | optimal 6% | optimal 4% | optimal 4% | 230 BF16 | $6.65 |
| Falcon3 3B | 3.0B | optimal 6% | optimal 4% | optimal 4% | 234 BF16 | $6.51 |
| Qwen 2.5 3B | 3.0B | optimal 6% | optimal 4% | optimal 4% | 235 BF16 | $6.50 |
| Qwen 2.5 Coder 3B | 3.0B | optimal 6% | optimal 4% | optimal 4% | 235 BF16 | $6.50 |
| StarCoder2 3B | 3.0B | optimal 6% | optimal 4% | optimal 3% | 235 BF16 | $6.50 |
| Qwen3-VL 2B | 2.1B | optimal 5% | optimal 4% | optimal 4% | 256 BF16 | $5.97 |
| Granite 3.1 2B | 2.0B | optimal 5% | optimal 4% | optimal 3% | 260 BF16 | $5.88 |
| Qwen3 1.7B | 1.7B | optimal 4% | optimal 4% | optimal 4% | 268 BF16 | $5.69 |
| SmolLM2 1.7B | 1.7B | optimal 5% | optimal 4% | optimal 4% | 267 BF16 | $5.71 |
| StableLM 2 1.6B 4K ctx | 1.6B | optimal 4% | optimal 4% | optimal 3% | 270 BF16 | $5.65 |
| Qwen 2 1.5B | 1.5B | optimal 4% | optimal 3% | optimal 3% | 276 BF16 | $5.55 |
| Qwen 2.5 1.5B | 1.5B | optimal 4% | optimal 3% | optimal 3% | 276 BF16 | $5.55 |
| Qwen 2.5 Coder 1.5B | 1.5B | optimal 4% | optimal 3% | optimal 3% | 276 BF16 | $5.55 |
| LLaMA 3.2 1B | 1.2B | optimal 4% | optimal 3% | optimal 3% | 284 BF16 | $5.38 |
| Falcon3 1B 4K ctx | 1.0B | optimal 3% | optimal 3% | optimal 3% | 292 BF16 | $5.24 |
| Gemma 3 1B | 1.0B | optimal 3% | optimal 3% | optimal 3% | 292 BF16 | $5.23 |
| Qwen3 0.6B | 0.6B | optimal 3% | optimal 3% | optimal 3% | 306 BF16 | $4.99 |
| Qwen 2 0.5B | 0.5B | optimal 3% | optimal 2% | optimal 3% | 312 BF16 | $4.90 |
| Qwen 2.5 0.5B | 0.5B | optimal 3% | optimal 2% | optimal 3% | 312 BF16 | $4.90 |
| Qwen 2.5 Coder 0.5B | 0.5B | optimal 3% | optimal 2% | optimal 3% | 312 BF16 | $4.90 |
| Kimi K3 | 2800.0B (104.0B active) | — | 32× GPUs | 16× GPUs | 55 FP8, 32× GPUs | $884.07 |
| DeepSeek V4-Pro | 1600.0B (49.0B active) | 32× GPUs | 16× GPUs | 8× GPUs | 101 FP8, 16× GPUs | $242.50 |
| Kimi K2 | 1000.0B (32.0B active) | 16× GPUs | 8× GPUs | 4× GPUs | 138 FP8, 8× GPUs | $88.50 |
| Kimi K2.5 | 1000.0B (32.0B active) | 16× GPUs | 8× GPUs | 4× GPUs | 138 FP8, 8× GPUs | $88.50 |
| GLM-5 744B | 744.0B (40.0B active) | 16× GPUs | 8× GPUs | 4× GPUs | 110 FP8, 8× GPUs | $111.31 |
| GLM-5.1 744B | 744.0B (40.0B active) | 16× GPUs | 8× GPUs | 4× GPUs | 110 FP8, 8× GPUs | $111.31 |
| DeepSeek R1 | 671.0B (37.0B active) | 16× GPUs | 8× GPUs | 4× GPUs | 114 FP8, 8× GPUs | $107.50 |
| DeepSeek R1-0528 | 671.0B (37.0B active) | 16× GPUs | 8× GPUs | 4× GPUs | 114 FP8, 8× GPUs | $107.50 |
| DeepSeek V3 | 671.0B (37.0B active) | 16× GPUs | 8× GPUs | 4× GPUs | 114 FP8, 8× GPUs | $107.50 |
| DeepSeek V3-0324 | 671.0B (37.0B active) | 16× GPUs | 8× GPUs | 4× GPUs | 114 FP8, 8× GPUs | $107.50 |
| DeepSeek V3.2 | 671.0B (37.0B active) | 16× GPUs | 8× GPUs | 4× GPUs | 114 FP8, 8× GPUs | $107.50 |
| MiniMax-M1 | 456.0B (45.9B active) | 8× GPUs | 4× GPUs | 2× GPUs | 93 FP8, 4× GPUs | $65.50 |
| MiniMax-Text-01 | 456.0B (45.9B active) | 8× GPUs | 4× GPUs | 2× GPUs | 93 FP8, 4× GPUs | $65.50 |
| LLaMA 3.1 405B | 405.0B | 8× GPUs | 4× GPUs | 2× GPUs | 30 FP8, 4× GPUs | $205.07 |
| Llama 4 Maverick (400B) | 400.0B (17.0B active) | 8× GPUs | 4× GPUs | 2× GPUs | 134 FP8, 4× GPUs | $45.71 |
| Qwen3.5 397B-A17B | 397.0B (17.0B active) | 8× GPUs | 4× GPUs | 2× GPUs | 186 FP8, 4× GPUs | $32.84 |
| Nemotron-4 340B 4K ctx | 340.0B | 4× GPUs | 4× GPUs | 2× GPUs | 35 FP8, 4× GPUs | $175.10 |
| Falcon 40B | 40.0B | — | — | — | — | — |
| Falcon 7B | 7.0B | — | — | — | — | — |
Frequently asked questions
What is the largest LLM that fits on a single GB200 NVL36?
At 8K context and batch 1, the largest models in our catalog that fit on a GB200 NVL36 (186 GB) are: Qwen3-Next 80B-A3B (80B) at BF16; Mixtral 8x22B (141B) at FP8; MiMo V2-Flash (309.8B) at 4-bit (AWQ). Larger models need tensor parallelism across several GPUs or lower-precision weights.
Can a GB200 NVL36 run LLaMA 3.3 70B?
at BF16: fits on one GB200 NVL36 (79% of VRAM, ~31 tok/s); at FP8: fits on one GB200 NVL36 (42% of VRAM, ~57 tok/s); at 4-bit (AWQ): fits on one GB200 NVL36 (25% of VRAM, ~97 tok/s). 8K context, batch 1, vLLM — the calculator sizes other contexts and concurrency.
Can a GB200 NVL36 serve DeepSeek R1?
at BF16: needs 16× GB200 NVL36; at FP8: needs 8× GB200 NVL36; at 4-bit (AWQ): needs 4× GB200 NVL36. 8K context, batch 1, vLLM — the calculator sizes other contexts and concurrency.
How fast is a GB200 NVL36 for LLM inference?
Single-stream decode for LLaMA 3.1 8B at FP8 is about 213 tokens/s on one GB200 NVL36; decode is bound by memory bandwidth (8000 GB/s), so larger models run proportionally slower. Batching raises total throughput far above the single-stream figure.
How much does it cost to rent a GB200 NVL36?
For on-demand rental our catalog estimate is $5.50/hr. The $/1M tokens column above uses that price.
Methodology: every cell comes from the same engine as our GPU Lookup — weights, KV cache (including MLA latent and hybrid / sliding-window layers), activations and framework overhead per model architecture, decode speed from a calibrated memory-bandwidth roofline. Reference scenario: 8K context (4K for models with a shorter native window), batch 1, vLLM, fp16 KV cache. Fit bands: optimal 50–70% of VRAM, good 70–80%, tight 80–95%. Engineering estimates, not a guarantee — validate against your serving framework before buying hardware.