GPUs for on-prem AI / NVIDIA
NVIDIA L4 24GB for LLMs
Low-profile 72W card for small-model inference (7B to 13B quantized) and dense edge servers; 300 GB/s bandwidth limits decode speed. NVIDIA lists 242/485 TFLOPS with sparsity, halved here. Price is an indicative 2026 street range for one unit, not a quote.
PCIe | PCIe Gen4 only | CUDA | indicative price $2,000-3,000
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 |
|---|---|---|---|---|---|
| Qwen3-32B | 32.8B | 1 | 9 | 4 | 13 |
| Olmo 3.1 32B Think | 32B | 1 | 10 | 4 | 15 |
| GLM-4.7-Flash | 31.2B / 3B | 1 | 67 | 2 | 41 |
| Gemma 4 31B IT | 30.7B | 1 | 10 | 4 | 14 |
| Qwen3-30B-A3B-Instruct-2507 | 30.5B / 3.3B | 1 | 60 | 2 | 30 |
| Granite 4.2 30B | 30B | 1 | 10 | 4 | 14 |
| NVIDIA Nemotron 3.5 Lightning 30B-A3B | 30B / 3B | 1 | 71 | 2 | 63 |
| Qwen3.8-27B | 27.8B | 1 | 11 | 2 | 10 |
| Gemma 3 27B IT | 27B | 1 | 11 | 2 | 10 |
| Gemma 4 26B A4B IT | 25.2B / 3.8B | 1 | 53 | 2 | 34 |
| Mistral Small 3.2 24B Instruct 2506 | 24B | 1 | 13 | 2 | 10 |
| Magistral Small 1.2 (2509) | 24B | 1 | 13 | 2 | 10 |
| Devstral Small 2 24B Instruct 2512 | 24B | 1 | 13 | 2 | 10 |
| gpt-oss-20b | 21B / 3.6B | 1 | 59 | 2 | 47 |
| Phi-4 | 14B | 1 | 21 | 2 | 14 |
| Gemma 4 12B IT | 11.95B | 1 | 25 | 1 | 11 |
| Qwen3.5-9B | 9.7B | 1 | 31 | 1 | 16 |
| Qwen3-VL-8B-Instruct | 8.8B | 1 | 33 | 2 | 21 |
| Granite 4.2 8B | 8B | 1 | 35 | 2 | 21 |
| Llama 3.1 8B Instruct | 8B | 1 | 36 | 1 | 13 |
| Qwen3.6-35B-A3B | 36B / 3B | 2 | 106 | 2 | 55 |
| Mistral Medium 3.5 128B | 128B | 4 | 8 | 8 | 8 |
| Qwen3.5-122B-A10B | 125B / 10B | 4 | 64 | 8 | 58 |
| Ling-3.0-flash | 124B / 5.1B | 4 | 106 | 8 | 97 |
| NVIDIA Nemotron 3 Super 120B-A12B | 120B / 12B | 4 | 51 | 8 | 47 |
| Mistral Small 4 119B A6B (2603) | 119B / 6.5B | 4 | 91 | 8 | 81 |
| gpt-oss-120b | 117B / 5.1B | 4 | 107 | 8 | 88 |
| Command A (03-2025) | 111B | 4 | 9 | 8 | 9 |
| Llama 4 Scout 17B-16E Instruct | 109B / 17B | 4 | 40 | 8 | 30 |
| Qwen3-Coder-Next | 79.7B / 3B | 4 | 139 | 8 | 110 |
| Llama 3.3 70B Instruct | 70B | 4 | 14 | 8 | 13 |
| Command A+ (05-2026) | 218B / 25B | 8 | 50 | 16 (multi-node) | 39 |
| Step-3.7-Flash | 198B / 11B | 8 | 87 | 16 (multi-node) | 59 |
| Qwen3.8-Flash-Next | 180B / 6B | 8 | 120 | 16 (multi-node) | 78 |
| MiniMax-M3 | 428B / 23B | 16 (multi-node) | 54 | 24 (multi-node) | 42 |
| Qwen3.5-397B-A17B | 403B / 17B | 16 (multi-node) | 62 | 24 (multi-node) | 52 |
| Llama 4 Maverick 17B-128E Instruct | 400B / 17B | 16 (multi-node) | 67 | 24 (multi-node) | 48 |
| Trinity Large Thinking | 398B / 13B | 16 (multi-node) | 67 | 24 (multi-node) | 51 |
| GLM-5.3-Flash | 320B / 18B | 16 (multi-node) | 69 | 16 (multi-node) | 49 |
| MiMo-V2.6-Flash-RL | 309B / 15B | 16 (multi-node) | 73 | 16 (multi-node) | 52 |
| Hy3 | 295B / 21B | 16 (multi-node) | 48 | 24 (multi-node) | 33 |
| Muse Spark Open 816B-A42B | 816B / 42B | 24 (multi-node) | 45 | 48 (multi-node) | 43 |
| DeepSeek-V4.1-Flash-1017 | 748B / 16B | 24 (multi-node) | 85 | 40 (multi-node) | 77 |
| GLM-5.3 | 744B / 40B | 24 (multi-node) | 42 | 40 (multi-node) | 37 |
| Mistral Large 3 675B Instruct 2512 | 675B / 41B | 24 (multi-node) | 46 | 40 (multi-node) | 41 |
| DeepSeek-V3.2 | 671B / 37B | 24 (multi-node) | 49 | 40 (multi-node) | 43 |
| MiMo-V2.6-Pro-RL | 1020B / 42B | 32 (multi-node) | 49 | 56 (multi-node) | 44 |
| Kimi-K2.6 | 1000B / 32B | 32 (multi-node) | 59 | 56 (multi-node) | 53 |
| Kimi-K2.7-Code | 1000B / 32B | 32 (multi-node) | 59 | 56 (multi-node) | 53 |
| DeepSeek-V4-Pro-0813 | 1600B / 49B | 48 (multi-node) | 54 | - | - |
| Kimi-K3 | 2800B / 104B | - | - | - | - |
| Qwen3.8-2.4T-A95B | 2446B / 95B | - | - | - | - |
Design a deployment on NVIDIA L4 24GB
Preloaded with this hardware. Add your ERP, documents and users.
Questions
What is the largest model that fits on one NVIDIA L4 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 L4 24GB for LLM inference?
Generation speed is mostly set by memory bandwidth: 300 GB/s here. Each token reads the active weights once, so a model with 8 GB of active weights tops out near 38 tokens per second per stream before efficiency losses; batching many users raises total throughput.
Is NVIDIA L4 24GB a good choice for on-prem AI?
Low-profile 72W card for small-model inference (7B to 13B quantized) and dense edge servers; 300 GB/s bandwidth limits decode speed. NVIDIA lists 242/485 TFLOPS with sparsity, halved here. Price is an indicative 2026 street range for one unit, not a quote.