GPUs for on-prem AI / NVIDIA
NVIDIA L40S 48GB for LLMs
Popular PCIe inference card with FP8; a 70B model needs two cards at 4-bit or four at 8-bit. No NVLink, so tensor parallel runs over PCIe. Price is an indicative 2026 street range for one unit, not a quote.
PCIe | PCIe Gen4 only | CUDA | indicative price $7,500-10K
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 | 13 | 4 | 19 |
| Qwen3.6-35B-A3B | 36B / 3B | 1 | 168 | 1 | 87 |
| Qwen3-32B | 32.8B | 1 | 26 | 2 | 20 |
| Olmo 3.1 32B Think | 32B | 1 | 27 | 1 | 13 |
| GLM-4.7-Flash | 31.2B / 3B | 1 | 163 | 1 | 66 |
| Gemma 4 31B IT | 30.7B | 1 | 28 | 2 | 21 |
| Qwen3-30B-A3B-Instruct-2507 | 30.5B / 3.3B | 1 | 147 | 1 | 48 |
| Granite 4.2 30B | 30B | 1 | 29 | 2 | 21 |
| NVIDIA Nemotron 3.5 Lightning 30B-A3B | 30B / 3B | 1 | 171 | 1 | 101 |
| Qwen3.8-27B | 27.8B | 1 | 31 | 1 | 17 |
| Gemma 3 27B IT | 27B | 1 | 32 | 1 | 16 |
| Gemma 4 26B A4B IT | 25.2B / 3.8B | 1 | 134 | 1 | 54 |
| Mistral Small 3.2 24B Instruct 2506 | 24B | 1 | 36 | 1 | 16 |
| Magistral Small 1.2 (2509) | 24B | 1 | 36 | 1 | 16 |
| Devstral Small 2 24B Instruct 2512 | 24B | 1 | 36 | 1 | 16 |
| gpt-oss-20b | 21B / 3.6B | 1 | 146 | 1 | 74 |
| Phi-4 | 14B | 1 | 57 | 1 | 22 |
| Gemma 4 12B IT | 11.95B | 1 | 67 | 1 | 30 |
| Qwen3.5-9B | 9.7B | 1 | 83 | 1 | 44 |
| Qwen3-VL-8B-Instruct | 8.8B | 1 | 86 | 1 | 33 |
| Granite 4.2 8B | 8B | 1 | 92 | 1 | 33 |
| Llama 3.1 8B Instruct | 8B | 1 | 94 | 1 | 36 |
| Mistral Medium 3.5 128B | 128B | 2 | 12 | 4 | 12 |
| Qwen3.5-122B-A10B | 125B / 10B | 2 | 95 | 4 | 85 |
| Ling-3.0-flash | 124B / 5.1B | 2 | 155 | 4 | 137 |
| NVIDIA Nemotron 3 Super 120B-A12B | 120B / 12B | 2 | 78 | 4 | 70 |
| Mistral Small 4 119B A6B (2603) | 119B / 6.5B | 2 | 133 | 4 | 116 |
| gpt-oss-120b | 117B / 5.1B | 2 | 156 | 4 | 125 |
| Command A (03-2025) | 111B | 2 | 14 | 4 | 14 |
| Llama 4 Scout 17B-16E Instruct | 109B / 17B | 2 | 60 | 4 | 45 |
| Qwen3-Coder-Next | 79.7B / 3B | 2 | 203 | 2 | 122 |
| Hy3 | 295B / 21B | 4 | 71 | 8 | 49 |
| Command A+ (05-2026) | 218B / 25B | 4 | 74 | 8 | 67 |
| Step-3.7-Flash | 198B / 11B | 4 | 124 | 8 | 104 |
| Qwen3.8-Flash-Next | 180B / 6B | 4 | 167 | 8 | 139 |
| MiniMax-M3 | 428B / 23B | 8 | 96 | 16 (multi-node) | 61 |
| Qwen3.5-397B-A17B | 403B / 17B | 8 | 112 | 16 (multi-node) | 72 |
| Llama 4 Maverick 17B-128E Instruct | 400B / 17B | 8 | 119 | 16 (multi-node) | 72 |
| Trinity Large Thinking | 398B / 13B | 8 | 121 | 16 (multi-node) | 71 |
| GLM-5.3-Flash | 320B / 18B | 8 | 121 | 8 | 86 |
| MiMo-V2.6-Flash-RL | 309B / 15B | 8 | 128 | 8 | 92 |
| MiMo-V2.6-Pro-RL | 1020B / 42B | 16 (multi-node) | 61 | 32 (multi-node) | 57 |
| Kimi-K2.6 | 1000B / 32B | 16 (multi-node) | 72 | 32 (multi-node) | 67 |
| Kimi-K2.7-Code | 1000B / 32B | 16 (multi-node) | 72 | 32 (multi-node) | 67 |
| Muse Spark Open 816B-A42B | 816B / 42B | 16 (multi-node) | 65 | 24 (multi-node) | 54 |
| DeepSeek-V4.1-Flash-1017 | 748B / 16B | 16 (multi-node) | 112 | 24 (multi-node) | 98 |
| GLM-5.3 | 744B / 40B | 16 (multi-node) | 58 | 24 (multi-node) | 49 |
| Mistral Large 3 675B Instruct 2512 | 675B / 41B | 16 (multi-node) | 66 | 24 (multi-node) | 56 |
| DeepSeek-V3.2 | 671B / 37B | 16 (multi-node) | 69 | 24 (multi-node) | 58 |
| DeepSeek-V4-Pro-0813 | 1600B / 49B | 24 (multi-node) | 66 | 40 (multi-node) | 58 |
| Kimi-K3 | 2800B / 104B | 40 (multi-node) | 43 | - | - |
| Qwen3.8-2.4T-A95B | 2446B / 95B | 40 (multi-node) | 45 | 64 (multi-node) | 43 |
Design a deployment on NVIDIA L40S 48GB
Preloaded with this hardware. Add your ERP, documents and users.
Questions
What is the largest model that fits on one NVIDIA L40S 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 L40S 48GB for LLM inference?
Generation speed is mostly set by memory bandwidth: 864 GB/s here. Each token reads the active weights once, so a model with 8 GB of active weights tops out near 108 tokens per second per stream before efficiency losses; batching many users raises total throughput.
Is NVIDIA L40S 48GB a good choice for on-prem AI?
Popular PCIe inference card with FP8; a 70B model needs two cards at 4-bit or four at 8-bit. No NVLink, so tensor parallel runs over PCIe. Price is an indicative 2026 street range for one unit, not a quote.