Mistral AI · Dense · GQA
Mistral Small 24B Instruct 2501：各量化与上下文下的显存
23.6B 参数、40 层、8 个 KV 头。换算成 GB 是这个结果。
架构参数
| Parameters | 23.6B | Vendor model card |
|---|---|---|
| Layers | 40 | config.json |
| Hidden size | 5120 | config.json |
| Attention heads | 32 | config.json |
| KV heads | 8 | config.json (GQA) |
| Head dim | 128 | config.json |
| Max context | 32768 tokens | config.json |
Source: mistralai/Mistral-Small-24B-Instruct-2501 config.json
各量化精度下的显存
Total includes 8K of KV cache and 1.5 GB runtime overhead.
| 量化 | Weights | KV @ 8K | 总需求 | 可运行于 |
|---|---|---|---|---|
| FP16 / BF16 2.0 B/param | 47.2 GB | 1.3 GB | 50.0 GB | Mac Studio M4 Max 128GB |
| INT8 / Q8_0 1.06 B/param | 25.0 GB | 1.3 GB | 27.9 GB | Mac Studio M4 Max 128GB |
| Q4_K_M (4-bit) 0.6 B/param | 14.2 GB | 1.3 GB | 17.0 GB | Mac Studio M4 Max 128GB |
| Q5_K_M (5-bit) 0.72 B/param | 17.0 GB | 1.3 GB | 19.8 GB | Mac Studio M4 Max 128GB |
各上下文长度下的 KV 缓存
Q4_K_M weights plus 1.5 GB overhead. KV cache at FP16.
| 上下文 | KV 缓存 | 总需求 | 可运行于 |
|---|---|---|---|
| 4k | 0.7 GB | 16.3 GB | Mac Studio M4 Max 128GB |
| 8k | 1.3 GB | 17.0 GB | Mac Studio M4 Max 128GB |
| 16k | 2.7 GB | 18.3 GB | Mac Studio M4 Max 128GB |
| 32k | 5.4 GB | 21.0 GB | Mac Studio M4 Max 128GB |
| 64k | 10.7 GB | 26.4 GB | Mac Studio M4 Max 128GB |
| 128k | 21.5 GB | 37.1 GB | Mac Studio M4 Max 128GB |
结论
At Q4_K_M the weights alone are about 14.2 GB. With 8K of context and 1.5 GB of runtime overhead it lands near 17.0 GB. It fits: RTX 4090, RTX 3090, RTX 5090, Mac Studio M4 Max 128GB.
所有吞吐数字均为基于内存带宽与公开实测的估算值，显存数字按模型架构与量化假设推算。购买硬件前请以你自己的运行环境为准。