Best LLMs to run on V100 SXM2 32GB
datacenter 32 GB HBM2 · 900 GB/s · 125 TFLOPS BF16 · updated 2026-10-23
Headline answer
Of 141 models in our catalog, 76 fit on a single V100 SXM2 32GB at FP8, 93 at 4-bit (AWQ) and 58 at BF16 (8K context, batch 1). Largest that fits: Mistral Nemo 12B at BF16; Devstral 24B at FP8; Kimi Linear 48B-A3B at 4-bit (AWQ). Rental: $0.33/hr catalog estimate.
Change context, concurrency or framework in the calculator → V100 SXM2 32GB full specs
Every model on V100 SXM2 32GB
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 |
|---|---|---|---|---|---|---|
| Kimi Linear 48B-A3B | 48.0B (3.0B active) | 8× GPUs | 2× GPUs | tight 95% | 172 4-bit | $0.17 |
| Mixtral 8x7B | 46.7B (12.9B active) | 8× GPUs | 2× GPUs | tight 93% | 66 4-bit | $0.45 |
| Qwen3.6 35B-A3B | 35.0B (3.0B active) | 4× GPUs | 2× GPUs | good 76% | 170 4-bit | $0.18 |
| CodeLlama 34B | 34.0B | 4× GPUs | 2× GPUs | good 76% | 29 4-bit | $1.03 |
| Yi-1.5 34B 4K ctx | 34.0B | 4× GPUs | 2× GPUs | good 75% | 29 4-bit | $1.04 |
| DeepSeek R1 Distill 32B | 32.0B | 4× GPUs | 2× GPUs | good 79% | 30 4-bit | $0.98 |
| Nemotron-Terminal-32B | 32.0B | 4× GPUs | 2× GPUs | good 79% | 30 4-bit | $0.98 |
| OLMo-2 32B 4K ctx | 32.0B | 4× GPUs | 2× GPUs | good 73% | 30 4-bit | $0.98 |
| Qwen 2.5 32B | 32.0B | 4× GPUs | 2× GPUs | good 79% | 30 4-bit | $0.98 |
| Qwen 2.5 Coder 32B | 32.0B | 4× GPUs | 2× GPUs | good 79% | 30 4-bit | $0.98 |
| Qwen3 32B | 32.0B | 4× GPUs | 2× GPUs | good 79% | 30 4-bit | $0.98 |
| Nemotron-3 Nano 30B-A3B | 31.6B (3.2B active) | 4× GPUs | 2× GPUs | optimal 68% | 168 4-bit | $0.18 |
| Muse Glimmer 30B | 30.0B | 4× GPUs | 2× GPUs | good 72% | 32 4-bit | $0.92 |
| Qwen3 30B-A3B | 30.0B (3.0B active) | 4× GPUs | 2× GPUs | optimal 67% | 169 4-bit | $0.18 |
| Gemma 2 27B | 27.0B | 4× GPUs | 2× GPUs | good 77% | 35 4-bit | $0.85 |
| Gemma 3 27B | 27.0B | 4× GPUs | 2× GPUs | tight 83% | 35 4-bit | $0.86 |
| Qwen3.6 27B | 27.0B | 4× GPUs | 2× GPUs | good 75% | 35 4-bit | $0.85 |
| Devstral 24B | 24.0B | 2× GPUs | tight 93% | optimal 62% | 21 FP8 | $1.42 |
| Devstral Small 2 24B | 24.0B | 2× GPUs | tight 93% | optimal 62% | 21 FP8 | $1.42 |
| Magistral Small 24B | 24.0B | 2× GPUs | tight 93% | optimal 62% | 21 FP8 | $1.42 |
| Mistral Small 3.1 24B | 24.0B | 2× GPUs | tight 93% | optimal 62% | 21 FP8 | $1.42 |
| Codestral 22B | 22.0B | 2× GPUs | tight 89% | optimal 56% | 23 FP8 | $1.31 |
| GPT OSS 20B | 21.0B (3.6B active) | 2× GPUs | tight 81% | optimal 54% | 103 FP8 | $0.29 |
| InternLM 2.5 20B | 20.0B | 2× GPUs | tight 82% | optimal 56% | 25 FP8 | $1.20 |
| DeepSeek Coder V2 Lite | 16.0B (2.4B active) | 2× GPUs | optimal 64% | optimal 42% | 134 FP8 | $0.22 |
| StarCoder2 15B | 16.0B | 2× GPUs | optimal 66% | optimal 44% | 31 FP8 | $0.97 |
| DeepSeek R1 Distill 14B | 14.8B | 2× GPUs | optimal 65% | optimal 49% | 33 FP8 | $0.91 |
| Qwen 2.5 14B | 14.8B | 2× GPUs | optimal 65% | optimal 49% | 33 FP8 | $0.91 |
| Qwen3 14B | 14.8B | 2× GPUs | optimal 64% | optimal 48% | 33 FP8 | $0.91 |
| Phi 3 Medium | 14.0B | 2× GPUs | optimal 63% | optimal 42% | 34 FP8 | $0.87 |
| Phi-4 14B | 14.0B | 2× GPUs | optimal 63% | optimal 46% | 34 FP8 | $0.87 |
| Qwen 2.5 Coder 14B | 14.0B | 2× GPUs | optimal 62% | optimal 48% | 34 FP8 | $0.87 |
| OLMo-2 13B 4K ctx | 13.7B | 2× GPUs | optimal 67% | optimal 50% | 34 FP8 | $0.88 |
| LLaMA 2 13B 4K ctx | 13.0B | 2× GPUs | optimal 65% | optimal 46% | 35 FP8 | $0.84 |
| Gemma 3 12B | 12.0B | 2× GPUs | optimal 61% | optimal 52% | 39 FP8 | $0.77 |
| Mistral Nemo 12B | 12.0B | tight 92% | optimal 55% | optimal 43% | 21 BF16 | $1.42 |
| Nemotron-Nano-12B-v2 | 12.0B | tight 94% | optimal 58% | optimal 45% | 21 BF16 | $1.42 |
| StableLM 2 12B 4K ctx | 12.0B | tight 90% | optimal 54% | optimal 40% | 21 BF16 | $1.42 |
| Falcon3 10B | 10.0B | good 79% | optimal 49% | optimal 37% | 25 BF16 | $1.20 |
| GLM-4 9B | 9.0B | good 70% | optimal 42% | optimal 34% | 28 BF16 | $1.08 |
| Gemma 2 9B | 9.0B | good 78% | optimal 50% | optimal 45% | 27 BF16 | $1.09 |
| Nemotron-Nano-9B-v2 | 9.0B | good 75% | optimal 47% | optimal 39% | 27 BF16 | $1.09 |
| Yi-1.5 9B 4K ctx | 9.0B | good 70% | optimal 43% | optimal 31% | 27 BF16 | $1.09 |
| InternLM3 8B | 8.8B | optimal 69% | optimal 42% | optimal 33% | 28 BF16 | $1.06 |
| Qwen3-VL 8B | 8.8B | good 71% | optimal 44% | optimal 36% | 28 BF16 | $1.06 |
| MiniCPM-V 4.5 | 8.7B | good 71% | optimal 44% | optimal 36% | 28 BF16 | $1.06 |
| Ministral-3 8B | 8.5B | optimal 69% | optimal 43% | optimal 35% | 29 BF16 | $1.03 |
| Qwen2.5-VL 7B | 8.3B | optimal 66% | optimal 41% | optimal 33% | 30 BF16 | $1.01 |
| DeepSeek R1 Distill 8B | 8.0B | optimal 66% | optimal 42% | optimal 34% | 30 BF16 | $0.98 |
| Granite 3.3 8B | 8.0B | optimal 67% | optimal 42% | optimal 32% | 30 BF16 | $0.98 |
| LLaMA 3.1 8B | 8.0B | optimal 66% | optimal 42% | optimal 34% | 30 BF16 | $0.98 |
| Ministral 8B | 8.0B | optimal 66% | optimal 42% | optimal 34% | 30 BF16 | $0.98 |
| Qwen3 8B | 8.0B | optimal 66% | optimal 42% | optimal 35% | 30 BF16 | $0.98 |
| Qwen 2.5 7B | 7.6B | optimal 62% | optimal 38% | optimal 32% | 32 BF16 | $0.93 |
| Falcon3 7B | 7.5B | optimal 62% | optimal 39% | optimal 32% | 32 BF16 | $0.92 |
| Phi-3 Small 7B | 7.4B | optimal 62% | optimal 40% | optimal 32% | 33 BF16 | $0.91 |
| InternLM 2.5 7B | 7.0B | optimal 60% | optimal 38% | optimal 31% | 34 BF16 | $0.87 |
| LLaMA 2 7B 4K ctx | 7.0B | optimal 63% | optimal 42% | optimal 32% | 34 BF16 | $0.88 |
| Mistral 7B | 7.0B | optimal 60% | optimal 38% | optimal 29% | 34 BF16 | $0.87 |
| OLMo-2 7B 4K ctx | 7.0B | optimal 63% | optimal 42% | optimal 35% | 34 BF16 | $0.88 |
| Qwen 2 7B | 7.0B | optimal 58% | optimal 36% | optimal 31% | 34 BF16 | $0.86 |
| Qwen 2.5 Coder 7B | 7.0B | optimal 58% | optimal 36% | optimal 31% | 34 BF16 | $0.86 |
| StarCoder2 7B | 7.0B | optimal 58% | optimal 37% | optimal 28% | 34 BF16 | $0.86 |
| Yi-1.5 6B 4K ctx | 6.0B | optimal 51% | optimal 33% | optimal 26% | 40 BF16 | $0.75 |
| Qwen3-VL 4B | 4.4B | optimal 44% | optimal 31% | optimal 27% | 51 BF16 | $0.58 |
| Gemma 3 4B | 4.3B | optimal 43% | optimal 30% | optimal 30% | 52 BF16 | $0.57 |
| Nemotron-Mini-4B 4K ctx | 4.0B | optimal 39% | optimal 27% | optimal 28% | 56 BF16 | $0.54 |
| Qwen3 4B | 4.0B | optimal 42% | optimal 29% | optimal 27% | 56 BF16 | $0.54 |
| Ministral 3B | 3.8B | optimal 38% | optimal 27% | optimal 25% | 58 BF16 | $0.51 |
| Phi 3.5 Mini | 3.8B | optimal 46% | optimal 35% | optimal 30% | 57 BF16 | $0.53 |
| Phi-4 Mini | 3.8B | optimal 40% | optimal 28% | optimal 28% | 58 BF16 | $0.51 |
| Qwen2.5-VL 3B | 3.8B | optimal 37% | optimal 26% | optimal 23% | 59 BF16 | $0.50 |
| Ministral-3 3B | 3.4B | optimal 37% | optimal 26% | optimal 25% | 64 BF16 | $0.47 |
| LLaMA 3.2 3B | 3.2B | optimal 36% | optimal 26% | optimal 24% | 66 BF16 | $0.45 |
| Falcon3 3B | 3.0B | optimal 34% | optimal 25% | optimal 24% | 70 BF16 | $0.42 |
| Qwen 2.5 3B | 3.0B | optimal 32% | optimal 23% | optimal 21% | 71 BF16 | $0.42 |
| Qwen 2.5 Coder 3B | 3.0B | optimal 32% | optimal 23% | optimal 21% | 71 BF16 | $0.42 |
| StarCoder2 3B | 3.0B | optimal 32% | optimal 23% | optimal 20% | 71 BF16 | $0.42 |
| Qwen3-VL 2B | 2.1B | optimal 29% | optimal 22% | optimal 22% | 90 BF16 | $0.33 |
| Granite 3.1 2B | 2.0B | optimal 27% | optimal 21% | optimal 19% | 95 BF16 | $0.31 |
| Qwen3 1.7B | 1.7B | optimal 26% | optimal 21% | optimal 21% | 106 BF16 | $0.28 |
| SmolLM2 1.7B | 1.7B | optimal 28% | optimal 23% | optimal 22% | 104 BF16 | $0.29 |
| StableLM 2 1.6B 4K ctx | 1.6B | optimal 25% | optimal 20% | optimal 20% | 109 BF16 | $0.27 |
| Qwen 2 1.5B | 1.5B | optimal 23% | optimal 18% | optimal 18% | 116 BF16 | $0.26 |
| Qwen 2.5 1.5B | 1.5B | optimal 23% | optimal 18% | optimal 18% | 116 BF16 | $0.26 |
| Qwen 2.5 Coder 1.5B | 1.5B | optimal 23% | optimal 18% | optimal 18% | 116 BF16 | $0.26 |
| LLaMA 3.2 1B | 1.2B | optimal 21% | optimal 18% | optimal 18% | 131 BF16 | $0.23 |
| Falcon3 1B 4K ctx | 1.0B | optimal 20% | optimal 17% | optimal 18% | 147 BF16 | $0.20 |
| Gemma 3 1B | 1.0B | optimal 20% | optimal 17% | optimal 17% | 148 BF16 | $0.20 |
| Qwen3 0.6B | 0.6B | optimal 19% | optimal 18% | optimal 18% | 186 BF16 | $0.16 |
| Qwen 2 0.5B | 0.5B | optimal 16% | optimal 15% | optimal 15% | 205 BF16 | $0.14 |
| Qwen 2.5 0.5B | 0.5B | optimal 16% | optimal 15% | optimal 15% | 205 BF16 | $0.14 |
| Qwen 2.5 Coder 0.5B | 0.5B | optimal 16% | optimal 15% | optimal 15% | 205 BF16 | $0.14 |
| LLaMA 3.1 405B | 405.0B | 32× GPUs | 24× GPUs | 16× GPUs | 6 FP8, 24× GPUs | $129.94 |
| Nemotron-4 340B 4K ctx | 340.0B | 32× GPUs | 16× GPUs | 8× GPUs | 7 FP8, 16× GPUs | $72.19 |
| MiMo V2-Flash | 309.8B (15.0B active) | — | — | 8× GPUs | 83 4-bit, 8× GPUs | $2.87 |
| DeepSeek V4-Flash | 284.0B (13.0B active) | — | — | 8× GPUs | 78 4-bit, 8× GPUs | $3.07 |
| Llama-3.3-Nemotron Ultra 253B | 253.0B | 24× GPUs | 16× GPUs | 8× GPUs | 9 FP8, 16× GPUs | $54.14 |
| DeepSeek Coder V2 | 236.0B (21.0B active) | — | — | 8× GPUs | 47 4-bit, 8× GPUs | $5.11 |
| Qwen3 235B-A22B | 235.0B (22.0B active) | — | — | 8× GPUs | 63 4-bit, 8× GPUs | $3.80 |
| MiniMax M2.5 | 230.0B (10.0B active) | — | — | 8× GPUs | 122 4-bit, 8× GPUs | $1.95 |
| MiniMax M2.7 | 230.0B (10.0B active) | — | — | 8× GPUs | 122 4-bit, 8× GPUs | $1.95 |
| Mixtral 8x22B | 141.0B (39.0B active) | — | 8× GPUs | 4× GPUs | 36 FP8, 8× GPUs | $6.67 |
| DBRX 132B | 132.0B (36.0B active) | — | 8× GPUs | 4× GPUs | 41 FP8, 8× GPUs | $5.87 |
| Devstral 2 123B | 123.0B | 16× GPUs | 8× GPUs | 4× GPUs | 18 FP8, 8× GPUs | $13.16 |
| Mistral Large 2 123B | 123.0B | 16× GPUs | 8× GPUs | 4× GPUs | 18 FP8, 8× GPUs | $13.16 |
| Nemotron-3 Super 120B-A12B | 120.0B (12.0B active) | — | 8× GPUs | 4× GPUs | 42 FP8, 8× GPUs | $5.70 |
| Mistral Small 4 119B | 119.0B (11.0B active) | — | 8× GPUs | 4× GPUs | 41 FP8, 8× GPUs | $5.82 |
| GPT OSS 120B | 117.0B (5.1B active) | — | 8× GPUs | 4× GPUs | 121 FP8, 8× GPUs | $1.97 |
| Llama 4 Scout (109B) | 109.0B (17.0B active) | — | 8× GPUs | 4× GPUs | 29 FP8, 8× GPUs | $8.19 |
| Command R+ 104B | 104.0B | 16× GPUs | 8× GPUs | 4× GPUs | 21 FP8, 8× GPUs | $11.24 |
| Qwen3-Next 80B-A3B | 80.0B (3.0B active) | 8× GPUs | 4× GPUs | 2× GPUs | 124 FP8, 4× GPUs | $0.96 |
| Qwen 2 72B | 72.0B | 8× GPUs | 4× GPUs | 2× GPUs | 19 FP8, 4× GPUs | $6.14 |
| Qwen 2.5 72B | 72.0B | 8× GPUs | 4× GPUs | 2× GPUs | 19 FP8, 4× GPUs | $6.14 |
| CodeLlama 70B | 70.0B | 8× GPUs | 4× GPUs | 2× GPUs | 20 FP8, 4× GPUs | $5.96 |
| DeepSeek R1 Distill 70B | 70.0B | 8× GPUs | 4× GPUs | 2× GPUs | 20 FP8, 4× GPUs | $5.96 |
| LLaMA 2 70B 4K ctx | 70.0B | 8× GPUs | 4× GPUs | 2× GPUs | 20 FP8, 4× GPUs | $5.96 |
| LLaMA 3.1 70B | 70.0B | 8× GPUs | 4× GPUs | 2× GPUs | 20 FP8, 4× GPUs | $5.96 |
| LLaMA 3.3 70B | 70.0B | 8× GPUs | 4× GPUs | 2× GPUs | 20 FP8, 4× GPUs | $5.96 |
| Llama-3.1-Nemotron-70B | 70.0B | 8× GPUs | 4× GPUs | 2× GPUs | 20 FP8, 4× GPUs | $5.96 |
| Nemotron-H 56B Base | 56.0B | 8× GPUs | 4× GPUs | 2× GPUs | 24 FP8, 4× GPUs | $4.86 |
| Llama-3.1-Nemotron-51B | 51.0B | 8× GPUs | 4× GPUs | 2× GPUs | 27 FP8, 4× GPUs | $4.44 |
| Llama-3.3-Nemotron Super 49B | 49.0B | 4× GPUs | 2× GPUs | 2× GPUs | 18 FP8, 2× GPUs | $3.38 |
| Command R 35B | 35.0B | 4× GPUs | 2× GPUs | 2× GPUs | 24 FP8, 2× GPUs | $2.52 |
| Kimi K3 | 2800.0B (104.0B active) | — | — | — | — | — |
| DeepSeek V4-Pro | 1600.0B (49.0B active) | — | — | — | — | — |
| Kimi K2 | 1000.0B (32.0B active) | — | — | — | — | — |
| Kimi K2.5 | 1000.0B (32.0B active) | — | — | — | — | — |
| GLM-5 744B | 744.0B (40.0B active) | — | — | — | — | — |
| GLM-5.1 744B | 744.0B (40.0B active) | — | — | — | — | — |
| DeepSeek R1 | 671.0B (37.0B active) | — | — | — | — | — |
| DeepSeek R1-0528 | 671.0B (37.0B active) | — | — | — | — | — |
| DeepSeek V3 | 671.0B (37.0B active) | — | — | — | — | — |
| DeepSeek V3-0324 | 671.0B (37.0B active) | — | — | — | — | — |
| DeepSeek V3.2 | 671.0B (37.0B active) | — | — | — | — | — |
| MiniMax-M1 | 456.0B (45.9B active) | — | — | — | — | — |
| MiniMax-Text-01 | 456.0B (45.9B active) | — | — | — | — | — |
| Llama 4 Maverick (400B) | 400.0B (17.0B active) | — | — | — | — | — |
| Qwen3.5 397B-A17B | 397.0B (17.0B active) | — | — | — | — | — |
| Falcon 40B | 40.0B | — | — | — | — | — |
| Falcon 7B | 7.0B | — | — | — | — | — |
Compare V100 SXM2 32GB
Frequently asked questions
What is the largest LLM that fits on a single V100 SXM2 32GB?
At 8K context and batch 1, the largest models in our catalog that fit on a V100 SXM2 32GB (32 GB) are: Mistral Nemo 12B (12B) at BF16; Devstral 24B (24B) at FP8; Kimi Linear 48B-A3B (48B) at 4-bit (AWQ). Larger models need tensor parallelism across several GPUs or lower-precision weights.
Can a V100 SXM2 32GB run LLaMA 3.3 70B?
at BF16: needs 8× V100 SXM2 32GB; at FP8: needs 4× V100 SXM2 32GB; at 4-bit (AWQ): needs 2× V100 SXM2 32GB. 8K context, batch 1, vLLM — the calculator sizes other contexts and concurrency.
Can a V100 SXM2 32GB serve DeepSeek R1?
at BF16: does not fit within one 8-GPU node; at FP8: does not fit within one 8-GPU node; at 4-bit (AWQ): does not fit within one 8-GPU node. 8K context, batch 1, vLLM — the calculator sizes other contexts and concurrency.
How fast is a V100 SXM2 32GB for LLM inference?
Single-stream decode for LLaMA 3.1 8B at FP8 is about 56 tokens/s on one V100 SXM2 32GB; decode is bound by memory bandwidth (900 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 V100 SXM2 32GB?
For on-demand rental our catalog estimate is $0.33/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.