GPUs for on-prem AI / NVIDIA
NVIDIA RTX 6000 Ada 48GB for LLMs
Ada workstation card with FP8 and no NVLink. FP8 is half NVIDIA's 1,457 TFLOPS sparse figure; BF16 is derived as half of dense FP8. Price is an indicative 2026 street range for one unit, not a quote.
workstation card | PCIe Gen4 only | CUDA | indicative price $6,500-8,000
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 |
|---|---|---|---|---|---|
| Llama 3.3 70B Instruct | 70B | 1 | 14 | 4 | 21 |
| Qwen3.6-35B-A3B | 36B / 3B | 1 | 182 | 1 | 96 |
| Qwen3-32B | 32.8B | 1 | 29 | 2 | 22 |
| Olmo 3.1 32B Think | 32B | 1 | 30 | 1 | 15 |
| GLM-4.7-Flash | 31.2B / 3B | 1 | 177 | 1 | 73 |
| Gemma 4 31B IT | 30.7B | 1 | 31 | 2 | 23 |
| Qwen3-30B-A3B-Instruct-2507 | 30.5B / 3.3B | 1 | 159 | 1 | 53 |
| Granite 4.2 30B | 30B | 1 | 32 | 2 | 24 |
| NVIDIA Nemotron 3.5 Lightning 30B-A3B | 30B / 3B | 1 | 185 | 1 | 111 |
| Qwen3.8-27B | 27.8B | 1 | 35 | 1 | 18 |
| Gemma 3 27B IT | 27B | 1 | 35 | 1 | 17 |
| Gemma 4 26B A4B IT | 25.2B / 3.8B | 1 | 145 | 1 | 60 |
| Mistral Small 3.2 24B Instruct 2506 | 24B | 1 | 39 | 1 | 18 |
| Magistral Small 1.2 (2509) | 24B | 1 | 39 | 1 | 18 |
| Devstral Small 2 24B Instruct 2512 | 24B | 1 | 39 | 1 | 18 |
| gpt-oss-20b | 21B / 3.6B | 1 | 159 | 1 | 81 |
| Phi-4 | 14B | 1 | 63 | 1 | 25 |
| Gemma 4 12B IT | 11.95B | 1 | 73 | 1 | 33 |
| Qwen3.5-9B | 9.7B | 1 | 90 | 1 | 48 |
| Qwen3-VL-8B-Instruct | 8.8B | 1 | 94 | 1 | 36 |
| Granite 4.2 8B | 8B | 1 | 101 | 1 | 37 |
| Llama 3.1 8B Instruct | 8B | 1 | 102 | 1 | 40 |
| Mistral Medium 3.5 128B | 128B | 2 | 14 | 4 | 13 |
| Qwen3.5-122B-A10B | 125B / 10B | 2 | 103 | 4 | 91 |
| Ling-3.0-flash | 124B / 5.1B | 2 | 166 | 4 | 145 |
| NVIDIA Nemotron 3 Super 120B-A12B | 120B / 12B | 2 | 84 | 4 | 74 |
| Mistral Small 4 119B A6B (2603) | 119B / 6.5B | 2 | 144 | 4 | 125 |
| gpt-oss-120b | 117B / 5.1B | 2 | 167 | 4 | 134 |
| Command A (03-2025) | 111B | 2 | 16 | 4 | 16 |
| Llama 4 Scout 17B-16E Instruct | 109B / 17B | 2 | 66 | 4 | 49 |
| Qwen3-Coder-Next | 79.7B / 3B | 2 | 213 | 2 | 131 |
| Hy3 | 295B / 21B | 4 | 76 | 8 (multi-node) | 44 |
| Command A+ (05-2026) | 218B / 25B | 4 | 80 | 8 (multi-node) | 65 |
| Step-3.7-Flash | 198B / 11B | 4 | 132 | 8 (multi-node) | 88 |
| Qwen3.8-Flash-Next | 180B / 6B | 4 | 174 | 8 (multi-node) | 107 |
| MiniMax-M3 | 428B / 23B | 8 (multi-node) | 78 | 16 (multi-node) | 64 |
| Qwen3.5-397B-A17B | 403B / 17B | 8 (multi-node) | 86 | 16 (multi-node) | 75 |
| Llama 4 Maverick 17B-128E Instruct | 400B / 17B | 8 (multi-node) | 96 | 16 (multi-node) | 75 |
| Trinity Large Thinking | 398B / 13B | 8 (multi-node) | 92 | 16 (multi-node) | 73 |
| GLM-5.3-Flash | 320B / 18B | 8 (multi-node) | 99 | 8 (multi-node) | 76 |
| MiMo-V2.6-Flash-RL | 309B / 15B | 8 (multi-node) | 102 | 8 (multi-node) | 79 |
| MiMo-V2.6-Pro-RL | 1020B / 42B | 16 (multi-node) | 63 | 32 (multi-node) | 59 |
| Kimi-K2.6 | 1000B / 32B | 16 (multi-node) | 74 | 32 (multi-node) | 69 |
| Kimi-K2.7-Code | 1000B / 32B | 16 (multi-node) | 74 | 32 (multi-node) | 69 |
| Muse Spark Open 816B-A42B | 816B / 42B | 16 (multi-node) | 67 | 24 (multi-node) | 56 |
| DeepSeek-V4.1-Flash-1017 | 748B / 16B | 16 (multi-node) | 115 | 24 (multi-node) | 101 |
| GLM-5.3 | 744B / 40B | 16 (multi-node) | 60 | 24 (multi-node) | 51 |
| Mistral Large 3 675B Instruct 2512 | 675B / 41B | 16 (multi-node) | 69 | 24 (multi-node) | 58 |
| DeepSeek-V3.2 | 671B / 37B | 16 (multi-node) | 71 | 24 (multi-node) | 60 |
| DeepSeek-V4-Pro-0813 | 1600B / 49B | 24 (multi-node) | 68 | 40 (multi-node) | 60 |
| Kimi-K3 | 2800B / 104B | 40 (multi-node) | 44 | - | - |
| Qwen3.8-2.4T-A95B | 2446B / 95B | 40 (multi-node) | 46 | 64 (multi-node) | 44 |
Design a deployment on NVIDIA RTX 6000 Ada 48GB
Preloaded with this hardware. Add your ERP, documents and users.
Questions
What is the largest model that fits on one NVIDIA RTX 6000 Ada 48GB?
Of the models we track, Llama 3.3 70B Instruct (70B) fits on a single 48 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 6000 Ada 48GB for LLM inference?
Generation speed is mostly set by memory bandwidth: 960 GB/s here. Each token reads the active weights once, so a model with 8 GB of active weights tops out near 120 tokens per second per stream before efficiency losses; batching many users raises total throughput.
Is NVIDIA RTX 6000 Ada 48GB a good choice for on-prem AI?
Ada workstation card with FP8 and no NVLink. FP8 is half NVIDIA's 1,457 TFLOPS sparse figure; BF16 is derived as half of dense FP8. Price is an indicative 2026 street range for one unit, not a quote.