GPUs for on-prem AI / NVIDIA
NVIDIA GeForce RTX 4090 24GB for LLMs
Fast for single-user 7B to 32B quantized models, but 24GB and no NVLink. Dense rates with FP32 accumulate per NVIDIA whitepaper (FP8 is 660 with FP16 accumulate). Discontinued; mostly used market. Price is an indicative 2026 street range for one unit, not a quote.
workstation card | PCIe Gen4 only | CUDA | indicative price $1,600-2,500
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 |
|---|---|---|---|---|---|
| Qwen3-32B | 32.8B | 1 | 30 | 4 | 39 |
| Olmo 3.1 32B Think | 32B | 1 | 31 | 4 | 44 |
| GLM-4.7-Flash | 31.2B / 3B | 1 | 183 | 2 | 112 |
| Gemma 4 31B IT | 30.7B | 1 | 32 | 4 | 41 |
| Qwen3-30B-A3B-Instruct-2507 | 30.5B / 3.3B | 1 | 165 | 2 | 86 |
| Granite 4.2 30B | 30B | 1 | 33 | 4 | 42 |
| NVIDIA Nemotron 3.5 Lightning 30B-A3B | 30B / 3B | 1 | 191 | 2 | 153 |
| Qwen3.8-27B | 27.8B | 1 | 36 | 2 | 32 |
| Gemma 3 27B IT | 27B | 1 | 37 | 2 | 31 |
| Gemma 4 26B A4B IT | 25.2B / 3.8B | 1 | 151 | 2 | 98 |
| Mistral Small 3.2 24B Instruct 2506 | 24B | 1 | 41 | 2 | 33 |
| Magistral Small 1.2 (2509) | 24B | 1 | 41 | 2 | 33 |
| Devstral Small 2 24B Instruct 2512 | 24B | 1 | 41 | 2 | 33 |
| gpt-oss-20b | 21B / 3.6B | 1 | 165 | 2 | 130 |
| Phi-4 | 14B | 1 | 66 | 2 | 43 |
| Gemma 4 12B IT | 11.95B | 1 | 76 | 1 | 35 |
| Qwen3.5-9B | 9.7B | 1 | 94 | 1 | 50 |
| Qwen3-VL-8B-Instruct | 8.8B | 1 | 98 | 2 | 62 |
| Granite 4.2 8B | 8B | 1 | 105 | 2 | 62 |
| Llama 3.1 8B Instruct | 8B | 1 | 107 | 1 | 42 |
| Qwen3.6-35B-A3B | 36B / 3B | 2 | 227 | 2 | 142 |
| Mistral Medium 3.5 128B | 128B | 4 | 25 | 8 (multi-node) | 21 |
| Qwen3.5-122B-A10B | 125B / 10B | 4 | 141 | 8 (multi-node) | 95 |
| Ling-3.0-flash | 124B / 5.1B | 4 | 198 | 8 (multi-node) | 125 |
| NVIDIA Nemotron 3 Super 120B-A12B | 120B / 12B | 4 | 107 | 8 (multi-node) | 65 |
| Mistral Small 4 119B A6B (2603) | 119B / 6.5B | 4 | 188 | 8 (multi-node) | 124 |
| gpt-oss-120b | 117B / 5.1B | 4 | 206 | 8 (multi-node) | 129 |
| Command A (03-2025) | 111B | 4 | 29 | 8 (multi-node) | 26 |
| Llama 4 Scout 17B-16E Instruct | 109B / 17B | 4 | 101 | 8 (multi-node) | 65 |
| Qwen3-Coder-Next | 79.7B / 3B | 4 | 220 | 8 (multi-node) | 122 |
| Llama 3.3 70B Instruct | 70B | 4 | 41 | 8 (multi-node) | 31 |
| Command A+ (05-2026) | 218B / 25B | 8 (multi-node) | 101 | 16 (multi-node) | 86 |
| Step-3.7-Flash | 198B / 11B | 8 (multi-node) | 116 | 16 (multi-node) | 98 |
| Qwen3.8-Flash-Next | 180B / 6B | 8 (multi-node) | 126 | 16 (multi-node) | 108 |
| MiniMax-M3 | 428B / 23B | 16 (multi-node) | 83 | 24 (multi-node) | 69 |
| Qwen3.5-397B-A17B | 403B / 17B | 16 (multi-node) | 88 | 24 (multi-node) | 76 |
| Llama 4 Maverick 17B-128E Instruct | 400B / 17B | 16 (multi-node) | 102 | 24 (multi-node) | 82 |
| Trinity Large Thinking | 398B / 13B | 16 (multi-node) | 91 | 24 (multi-node) | 76 |
| GLM-5.3-Flash | 320B / 18B | 16 (multi-node) | 105 | 16 (multi-node) | 89 |
| MiMo-V2.6-Flash-RL | 309B / 15B | 16 (multi-node) | 105 | 16 (multi-node) | 89 |
| Hy3 | 295B / 21B | 16 (multi-node) | 69 | 24 (multi-node) | 54 |
| Muse Spark Open 816B-A42B | 816B / 42B | 24 (multi-node) | 71 | 48 (multi-node) | 65 |
| DeepSeek-V4.1-Flash-1017 | 748B / 16B | 24 (multi-node) | 114 | 40 (multi-node) | 103 |
| GLM-5.3 | 744B / 40B | 24 (multi-node) | 61 | 40 (multi-node) | 54 |
| Mistral Large 3 675B Instruct 2512 | 675B / 41B | 24 (multi-node) | 72 | 40 (multi-node) | 64 |
| DeepSeek-V3.2 | 671B / 37B | 24 (multi-node) | 74 | 40 (multi-node) | 66 |
| MiMo-V2.6-Pro-RL | 1020B / 42B | 32 (multi-node) | 69 | 56 (multi-node) | 62 |
| Kimi-K2.6 | 1000B / 32B | 32 (multi-node) | 80 | 56 (multi-node) | 71 |
| Kimi-K2.7-Code | 1000B / 32B | 32 (multi-node) | 80 | 56 (multi-node) | 71 |
| DeepSeek-V4-Pro-0813 | 1600B / 49B | 48 (multi-node) | 72 | - | - |
| Kimi-K3 | 2800B / 104B | - | - | - | - |
| Qwen3.8-2.4T-A95B | 2446B / 95B | - | - | - | - |
Design a deployment on NVIDIA GeForce RTX 4090 24GB
Preloaded with this hardware. Add your ERP, documents and users.
Questions
What is the largest model that fits on one NVIDIA GeForce RTX 4090 24GB?
Of the models we track, Qwen3-32B (32.8B) fits on a single 24 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 GeForce RTX 4090 24GB for LLM inference?
Generation speed is mostly set by memory bandwidth: 1008 GB/s here. Each token reads the active weights once, so a model with 8 GB of active weights tops out near 126 tokens per second per stream before efficiency losses; batching many users raises total throughput.
Is NVIDIA GeForce RTX 4090 24GB a good choice for on-prem AI?
Fast for single-user 7B to 32B quantized models, but 24GB and no NVLink. Dense rates with FP32 accumulate per NVIDIA whitepaper (FP8 is 660 with FP16 accumulate). Discontinued; mostly used market. Price is an indicative 2026 street range for one unit, not a quote.