GPUs for on-prem AI / NVIDIA
NVIDIA H200 SXM 141GB for LLMs
Same compute as H100 SXM with 141GB and 4.8 TB/s, so larger models and longer contexts per GPU and faster decode. NVIDIA lists 1,979/3,958 TFLOPS with sparsity, halved here. Price is an indicative 2026 street range for one unit, not a quote.
SXM5 | NVLink 900 GB/s | CUDA | indicative price $30K-40K
Which models fit, and how many cards
Planning estimates. 4-bit column: one user at 8K context. FP8 column: 10 users at once at 16K context.
| Model | Size | 4-bit: cards | 4-bit: tok/s | FP8, 10 users: cards | FP8: tok/s per user |
|---|---|---|---|---|---|
| Command A+ (05-2026) | 218B / 25B | 1 | 123 | 2 | 106 |
| Step-3.7-Flash | 198B / 11B | 1 | 226 | 2 | 171 |
| Qwen3.8-Flash-Next | 180B / 6B | 1 | 325 | 2 | 234 |
| Mistral Medium 3.5 128B | 128B | 1 | 38 | 2 | 32 |
| Qwen3.5-122B-A10B | 125B / 10B | 1 | 243 | 2 | 189 |
| Ling-3.0-flash | 124B / 5.1B | 1 | 353 | 1 | 249 |
| NVIDIA Nemotron 3 Super 120B-A12B | 120B / 12B | 1 | 216 | 1 | 139 |
| Mistral Small 4 119B A6B (2603) | 119B / 6.5B | 1 | 312 | 1 | 202 |
| gpt-oss-120b | 117B / 5.1B | 1 | 350 | 1 | 218 |
| Command A (03-2025) | 111B | 1 | 43 | 1 | 24 |
| Llama 4 Scout 17B-16E Instruct | 109B / 17B | 1 | 164 | 1 | 76 |
| Qwen3-Coder-Next | 79.7B / 3B | 1 | 434 | 1 | 297 |
| Llama 3.3 70B Instruct | 70B | 1 | 65 | 1 | 33 |
| Qwen3.6-35B-A3B | 36B / 3B | 1 | 435 | 1 | 304 |
| Qwen3-32B | 32.8B | 1 | 124 | 1 | 60 |
| Olmo 3.1 32B Think | 32B | 1 | 126 | 1 | 67 |
| GLM-4.7-Flash | 31.2B / 3B | 1 | 429 | 1 | 253 |
| Gemma 4 31B IT | 30.7B | 1 | 129 | 1 | 62 |
| Qwen3-30B-A3B-Instruct-2507 | 30.5B / 3.3B | 1 | 407 | 1 | 202 |
| Granite 4.2 30B | 30B | 1 | 133 | 1 | 63 |
| NVIDIA Nemotron 3.5 Lightning 30B-A3B | 30B / 3B | 1 | 438 | 1 | 332 |
| Qwen3.8-27B | 27.8B | 1 | 144 | 1 | 83 |
| Gemma 3 27B IT | 27B | 1 | 145 | 1 | 78 |
| Gemma 4 26B A4B IT | 25.2B / 3.8B | 1 | 388 | 1 | 220 |
| Mistral Small 3.2 24B Instruct 2506 | 24B | 1 | 159 | 1 | 82 |
| Magistral Small 1.2 (2509) | 24B | 1 | 159 | 1 | 82 |
| Devstral Small 2 24B Instruct 2512 | 24B | 1 | 159 | 1 | 82 |
| gpt-oss-20b | 21B / 3.6B | 1 | 407 | 1 | 274 |
| Phi-4 | 14B | 1 | 228 | 1 | 107 |
| Gemma 4 12B IT | 11.95B | 1 | 254 | 1 | 139 |
| Qwen3.5-9B | 9.7B | 1 | 293 | 1 | 187 |
| Qwen3-VL-8B-Instruct | 8.8B | 1 | 301 | 1 | 149 |
| Granite 4.2 8B | 8B | 1 | 315 | 1 | 150 |
| Llama 3.1 8B Instruct | 8B | 1 | 317 | 1 | 161 |
| MiniMax-M3 | 428B / 23B | 2 | 164 | 4 | 129 |
| Qwen3.5-397B-A17B | 403B / 17B | 2 | 194 | 4 | 156 |
| Llama 4 Maverick 17B-128E Instruct | 400B / 17B | 2 | 199 | 4 | 146 |
| Trinity Large Thinking | 398B / 13B | 2 | 213 | 4 | 153 |
| GLM-5.3-Flash | 320B / 18B | 2 | 200 | 4 | 172 |
| MiMo-V2.6-Flash-RL | 309B / 15B | 2 | 216 | 4 | 178 |
| Hy3 | 295B / 21B | 2 | 159 | 4 | 106 |
| Muse Spark Open 816B-A42B | 816B / 42B | 4 | 138 | 8 | 116 |
| DeepSeek-V4.1-Flash-1017 | 748B / 16B | 4 | 231 | 8 | 197 |
| GLM-5.3 | 744B / 40B | 4 | 130 | 8 | 107 |
| Mistral Large 3 675B Instruct 2512 | 675B / 41B | 4 | 141 | 8 | 120 |
| DeepSeek-V3.2 | 671B / 37B | 4 | 147 | 8 | 124 |
| DeepSeek-V4-Pro-0813 | 1600B / 49B | 8 | 140 | 16 (multi-node) | 84 |
| MiMo-V2.6-Pro-RL | 1020B / 42B | 8 | 135 | 16 (multi-node) | 79 |
| Kimi-K2.6 | 1000B / 32B | 8 | 155 | 16 (multi-node) | 91 |
| Kimi-K2.7-Code | 1000B / 32B | 8 | 155 | 16 (multi-node) | 91 |
| Kimi-K3 | 2800B / 104B | 16 (multi-node) | 62 | 24 (multi-node) | 54 |
| Qwen3.8-2.4T-A95B | 2446B / 95B | 16 (multi-node) | 63 | 24 (multi-node) | 55 |
Design a deployment on NVIDIA H200 SXM 141GB
Preloaded with this hardware. Add your ERP, documents and users.
Questions
What is the largest model that fits on one NVIDIA H200 SXM 141GB?
Of the models we track, Command A+ (05-2026) (218B total, 25B active) fits on a single 141 GB card in 4-bit with one user at 8K context. More users or longer context need memory for the KV cache, so leave headroom.
How fast is NVIDIA H200 SXM 141GB for LLM inference?
Generation speed is mostly set by memory bandwidth: 4800 GB/s here. Each token reads the active weights once, so a model with 8 GB of active weights tops out near 600 tokens per second per stream before efficiency losses; batching many users raises total throughput.
Is NVIDIA H200 SXM 141GB a good choice for on-prem AI?
Same compute as H100 SXM with 141GB and 4.8 TB/s, so larger models and longer contexts per GPU and faster decode. NVIDIA lists 1,979/3,958 TFLOPS with sparsity, halved here. Price is an indicative 2026 street range for one unit, not a quote.