GPUs for on-prem AI / NVIDIA
NVIDIA A10 24GB for LLMs
Ampere inference card for 7B to 13B models; no FP8, so quantized serving relies on INT8/INT4 weight-only kernels. NVIDIA lists 125 TFLOPS BF16 dense (250 with sparsity). Price is an indicative 2026 street range for one unit, not a quote.
PCIe | PCIe Gen4 only | CUDA | indicative price $1,500-3,000
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 | 18 | 4 | 25 |
| Olmo 3.1 32B Think | 32B | 1 | 19 | 4 | 28 |
| GLM-4.7-Flash | 31.2B / 3B | 1 | 123 | 2 | 75 |
| Gemma 4 31B IT | 30.7B | 1 | 19 | 4 | 26 |
| Qwen3-30B-A3B-Instruct-2507 | 30.5B / 3.3B | 1 | 109 | 2 | 56 |
| Granite 4.2 30B | 30B | 1 | 20 | 4 | 27 |
| NVIDIA Nemotron 3.5 Lightning 30B-A3B | 30B / 3B | 1 | 129 | 2 | 108 |
| Qwen3.8-27B | 27.8B | 1 | 22 | 2 | 20 |
| Gemma 3 27B IT | 27B | 1 | 22 | 2 | 19 |
| Gemma 4 26B A4B IT | 25.2B / 3.8B | 1 | 99 | 2 | 64 |
| Mistral Small 3.2 24B Instruct 2506 | 24B | 1 | 25 | 2 | 20 |
| Magistral Small 1.2 (2509) | 24B | 1 | 25 | 2 | 20 |
| Devstral Small 2 24B Instruct 2512 | 24B | 1 | 25 | 2 | 20 |
| gpt-oss-20b | 21B / 3.6B | 1 | 109 | 2 | 87 |
| Phi-4 | 14B | 1 | 41 | 2 | 27 |
| Gemma 4 12B IT | 11.95B | 1 | 48 | 1 | 21 |
| Qwen3.5-9B | 9.7B | 1 | 60 | 1 | 31 |
| Qwen3-VL-8B-Instruct | 8.8B | 1 | 62 | 2 | 39 |
| Granite 4.2 8B | 8B | 1 | 67 | 2 | 39 |
| Llama 3.1 8B Instruct | 8B | 1 | 68 | 1 | 26 |
| Qwen3.6-35B-A3B | 36B / 3B | 2 | 171 | 2 | 98 |
| Mistral Medium 3.5 128B | 128B | 4 | 16 | 8 (multi-node) | 14 |
| Qwen3.5-122B-A10B | 125B / 10B | 4 | 104 | 8 (multi-node) | 76 |
| Ling-3.0-flash | 124B / 5.1B | 4 | 158 | 8 (multi-node) | 108 |
| NVIDIA Nemotron 3 Super 120B-A12B | 120B / 12B | 4 | 81 | 8 (multi-node) | 55 |
| Mistral Small 4 119B A6B (2603) | 119B / 6.5B | 4 | 144 | 8 (multi-node) | 102 |
| gpt-oss-120b | 117B / 5.1B | 4 | 163 | 8 (multi-node) | 107 |
| Command A (03-2025) | 111B | 4 | 18 | 8 (multi-node) | 17 |
| Llama 4 Scout 17B-16E Instruct | 109B / 17B | 4 | 70 | 8 (multi-node) | 47 |
| Qwen3-Coder-Next | 79.7B / 3B | 4 | 188 | 8 (multi-node) | 110 |
| Llama 3.3 70B Instruct | 70B | 4 | 27 | 8 (multi-node) | 21 |
| Command A+ (05-2026) | 218B / 25B | 8 (multi-node) | 75 | 16 (multi-node) | 63 |
| Step-3.7-Flash | 198B / 11B | 8 (multi-node) | 99 | 16 (multi-node) | 82 |
| Qwen3.8-Flash-Next | 180B / 6B | 8 (multi-node) | 114 | 16 (multi-node) | 98 |
| MiniMax-M3 | 428B / 23B | 16 (multi-node) | 72 | 24 (multi-node) | 58 |
| Qwen3.5-397B-A17B | 403B / 17B | 16 (multi-node) | 79 | 24 (multi-node) | 67 |
| Llama 4 Maverick 17B-128E Instruct | 400B / 17B | 16 (multi-node) | 89 | 24 (multi-node) | 68 |
| Trinity Large Thinking | 398B / 13B | 16 (multi-node) | 83 | 24 (multi-node) | 66 |
| GLM-5.3-Flash | 320B / 18B | 16 (multi-node) | 91 | 16 (multi-node) | 72 |
| MiMo-V2.6-Flash-RL | 309B / 15B | 16 (multi-node) | 93 | 16 (multi-node) | 74 |
| Hy3 | 295B / 21B | 16 (multi-node) | 61 | 24 (multi-node) | 46 |
| Muse Spark Open 816B-A42B | 816B / 42B | 24 (multi-node) | 61 | 48 (multi-node) | 57 |
| DeepSeek-V4.1-Flash-1017 | 748B / 16B | 24 (multi-node) | 104 | 40 (multi-node) | 94 |
| GLM-5.3 | 744B / 40B | 24 (multi-node) | 54 | 40 (multi-node) | 48 |
| Mistral Large 3 675B Instruct 2512 | 675B / 41B | 24 (multi-node) | 62 | 40 (multi-node) | 55 |
| DeepSeek-V3.2 | 671B / 37B | 24 (multi-node) | 64 | 40 (multi-node) | 57 |
| MiMo-V2.6-Pro-RL | 1020B / 42B | 32 (multi-node) | 62 | 56 (multi-node) | 56 |
| Kimi-K2.6 | 1000B / 32B | 32 (multi-node) | 72 | 56 (multi-node) | 65 |
| Kimi-K2.7-Code | 1000B / 32B | 32 (multi-node) | 72 | 56 (multi-node) | 65 |
| DeepSeek-V4-Pro-0813 | 1600B / 49B | 48 (multi-node) | 66 | - | - |
| Kimi-K3 | 2800B / 104B | - | - | - | - |
| Qwen3.8-2.4T-A95B | 2446B / 95B | - | - | - | - |
Design a deployment on NVIDIA A10 24GB
Preloaded with this hardware. Add your ERP, documents and users.
Questions
What is the largest model that fits on one NVIDIA A10 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 A10 24GB for LLM inference?
Generation speed is mostly set by memory bandwidth: 600 GB/s here. Each token reads the active weights once, so a model with 8 GB of active weights tops out near 75 tokens per second per stream before efficiency losses; batching many users raises total throughput.
Is NVIDIA A10 24GB a good choice for on-prem AI?
Ampere inference card for 7B to 13B models; no FP8, so quantized serving relies on INT8/INT4 weight-only kernels. NVIDIA lists 125 TFLOPS BF16 dense (250 with sparsity). Price is an indicative 2026 street range for one unit, not a quote.