GPUs for on-prem AI / NVIDIA
NVIDIA RTX PRO 6000 Blackwell Server Edition 96GB for LLMs
Passive data center version (up to 600W configurable) for 8-GPU PCIe servers; strong inference per dollar with FP4 but no NVLink. NVIDIA lists 4 PF FP4, 2 PF FP8, 1 PF BF16 with sparsity, halved here. Price is an indicative 2026 street range for one unit, not a quote.
PCIe | PCIe Gen5 only | CUDA | indicative price $8,500-11K
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 |
|---|---|---|---|---|---|
| Mistral Medium 3.5 128B | 128B | 1 | 13 | 2 | 12 |
| Qwen3.5-122B-A10B | 125B / 10B | 1 | 107 | 2 | 91 |
| Ling-3.0-flash | 124B / 5.1B | 1 | 182 | 2 | 153 |
| NVIDIA Nemotron 3 Super 120B-A12B | 120B / 12B | 1 | 92 | 2 | 78 |
| Mistral Small 4 119B A6B (2603) | 119B / 6.5B | 1 | 151 | 2 | 125 |
| gpt-oss-120b | 117B / 5.1B | 1 | 179 | 2 | 135 |
| Command A (03-2025) | 111B | 1 | 15 | 2 | 14 |
| Llama 4 Scout 17B-16E Instruct | 109B / 17B | 1 | 65 | 2 | 46 |
| Qwen3-Coder-Next | 79.7B / 3B | 1 | 256 | 1 | 140 |
| Llama 3.3 70B Instruct | 70B | 1 | 23 | 2 | 19 |
| Qwen3.6-35B-A3B | 36B / 3B | 1 | 256 | 1 | 145 |
| Qwen3-32B | 32.8B | 1 | 47 | 1 | 21 |
| Olmo 3.1 32B Think | 32B | 1 | 48 | 1 | 24 |
| GLM-4.7-Flash | 31.2B / 3B | 1 | 250 | 1 | 113 |
| Gemma 4 31B IT | 30.7B | 1 | 49 | 1 | 22 |
| Qwen3-30B-A3B-Instruct-2507 | 30.5B / 3.3B | 1 | 229 | 1 | 84 |
| Granite 4.2 30B | 30B | 1 | 51 | 1 | 23 |
| NVIDIA Nemotron 3.5 Lightning 30B-A3B | 30B / 3B | 1 | 259 | 1 | 166 |
| Qwen3.8-27B | 27.8B | 1 | 56 | 1 | 30 |
| Gemma 3 27B IT | 27B | 1 | 57 | 1 | 28 |
| Gemma 4 26B A4B IT | 25.2B / 3.8B | 1 | 211 | 1 | 94 |
| Mistral Small 3.2 24B Instruct 2506 | 24B | 1 | 63 | 1 | 30 |
| Magistral Small 1.2 (2509) | 24B | 1 | 63 | 1 | 30 |
| Devstral Small 2 24B Instruct 2512 | 24B | 1 | 63 | 1 | 30 |
| gpt-oss-20b | 21B / 3.6B | 1 | 228 | 1 | 125 |
| Phi-4 | 14B | 1 | 99 | 1 | 40 |
| Gemma 4 12B IT | 11.95B | 1 | 114 | 1 | 54 |
| Qwen3.5-9B | 9.7B | 1 | 138 | 1 | 77 |
| Qwen3-VL-8B-Instruct | 8.8B | 1 | 143 | 1 | 58 |
| Granite 4.2 8B | 8B | 1 | 153 | 1 | 59 |
| Llama 3.1 8B Instruct | 8B | 1 | 154 | 1 | 64 |
| Hy3 | 295B / 21B | 2 | 78 | 4 | 53 |
| Command A+ (05-2026) | 218B / 25B | 2 | 76 | 4 | 69 |
| Step-3.7-Flash | 198B / 11B | 2 | 138 | 4 | 114 |
| Qwen3.8-Flash-Next | 180B / 6B | 2 | 196 | 4 | 159 |
| MiniMax-M3 | 428B / 23B | 4 | 108 | 8 | 89 |
| Qwen3.5-397B-A17B | 403B / 17B | 4 | 128 | 8 | 110 |
| Llama 4 Maverick 17B-128E Instruct | 400B / 17B | 4 | 133 | 8 | 102 |
| Trinity Large Thinking | 398B / 13B | 4 | 141 | 8 | 107 |
| GLM-5.3-Flash | 320B / 18B | 4 | 134 | 4 | 92 |
| MiMo-V2.6-Flash-RL | 309B / 15B | 4 | 145 | 4 | 99 |
| MiMo-V2.6-Pro-RL | 1020B / 42B | 8 | 92 | 16 (multi-node) | 60 |
| Kimi-K2.6 | 1000B / 32B | 8 | 110 | 16 (multi-node) | 71 |
| Kimi-K2.7-Code | 1000B / 32B | 8 | 110 | 16 (multi-node) | 71 |
| Muse Spark Open 816B-A42B | 816B / 42B | 8 | 96 | 16 (multi-node) | 63 |
| DeepSeek-V4.1-Flash-1017 | 748B / 16B | 8 | 171 | 16 (multi-node) | 112 |
| GLM-5.3 | 744B / 40B | 8 | 90 | 16 (multi-node) | 57 |
| Mistral Large 3 675B Instruct 2512 | 675B / 41B | 8 | 98 | 16 (multi-node) | 65 |
| DeepSeek-V3.2 | 671B / 37B | 8 | 103 | 8 | 72 |
| DeepSeek-V4-Pro-0813 | 1600B / 49B | 16 (multi-node) | 76 | 24 (multi-node) | 65 |
| Kimi-K3 | 2800B / 104B | 24 (multi-node) | 48 | 40 (multi-node) | 44 |
| Qwen3.8-2.4T-A95B | 2446B / 95B | 24 (multi-node) | 50 | 32 (multi-node) | 46 |
Design a deployment on NVIDIA RTX PRO 6000 Blackwell Server Edition 96GB
Preloaded with this hardware. Add your ERP, documents and users.
Questions
What is the largest model that fits on one NVIDIA RTX PRO 6000 Blackwell Server Edition 96GB?
Of the models we track, Mistral Medium 3.5 128B (128B) fits on a single 96 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 RTX PRO 6000 Blackwell Server Edition 96GB for LLM inference?
Generation speed is mostly set by memory bandwidth: 1597 GB/s here. Each token reads the active weights once, so a model with 8 GB of active weights tops out near 200 tokens per second per stream before efficiency losses; batching many users raises total throughput.
Is NVIDIA RTX PRO 6000 Blackwell Server Edition 96GB a good choice for on-prem AI?
Passive data center version (up to 600W configurable) for 8-GPU PCIe servers; strong inference per dollar with FP4 but no NVLink. NVIDIA lists 4 PF FP4, 2 PF FP8, 1 PF BF16 with sparsity, halved here. Price is an indicative 2026 street range for one unit, not a quote.