Mistral AI123BQ4_K_M 下约 75 GB 显存

Devstral-2 123B — 显存、速度与本地部署

作者: Jakub Rusinowski · 最后更新:

The flagship coding agent, and dense — Mistral deliberately did not build this one sparse, so all 123B stream per token and throughput scales accordingly. Scores 71.6% on SWE-bench Verified, the best open-source coding-agent result at release. Requires about 75 GB at Q4_K_M, so a dual-GPU workstation. Apache 2.0.

Devstral-2 123B 在 Q4_K_M 下约需 75 GB 显存——量化权重加框架开销,不含 KV 缓存。在 Apple Silicon 上,这部分来自统一内存。

逻辑87
创意78
编程95

按量化级别的显存与速度

计算基准:NVIDIA RTX 4090 (24 GB)。仅含权重与开销:该模型架构未公开,因此未计入 KV 缓存。

量化显存速度(估算)适配
Q2_K
2.63 bpw
41.2 GB~3 tok/s需卸载
Q3_K_M
3.41 bpw
53.2 GB~2 tok/s需卸载
Q4_K_M
4.83 bpw
75.1 GB—放不下
Q5_K_M
5.67 bpw
88 GB—放不下
Q6_K
6.56 bpw
101.7 GB—放不下
Q8_0
8.50 bpw
131.5 GB—放不下
F16
16.00 bpw
246.8 GB—放不下

黑色标记 = NVIDIA RTX 4090 (24 GB) 上的可用显存。 估算来自内存带宽屋顶线模型,详见 方法说明页. Devstral-2 123B 显存计算器 →

运行 Devstral-2 123B

目录中能运行 Devstral-2 123B 的最便宜 GPU 是 AMD Ryzen AI Max+ 395 (128 GB).

联盟营销声明: 本页部分链接为联盟推广链接——如果你通过它们购买,LLM Configurator 可能会获得佣金,而你无需支付任何额外费用。作为亚马逊联盟成员(Amazon Associate),LLM Configurator 会从符合条件的购买中获得收益。
Ryzen AI Max+ 395 Laptop (Strix Halo, up to 128GB)
128 GB VRAM · 120 W board power
2026年价格波动较大——请以当前商品页价格为准。

如何运行 Devstral-2 123B

安装 Ollama,然后运行:

ollama run devstral:123b
Hugging Face 上的权重: mistralai/Devstral-2-123B-Instruct-2512 ↗

规格

Preview — The model is released, but these specs are thin or rest on a single source. Individual fields may be wrong.

Preview — The model is released, but these specs are thin or rest on a single source. Individual fields may be wrong.

参数量
123 Billion (dense)
上下文窗口
262,144
架构
Dense Transformer
提供商
Mistral AI
许可证
Apache 2.0
规格量化
Q4_K_M
系统内存
128 GB
记录更新于
2026-04-05
许可证Apache-2.0允许商业使用

Commercial use permitted. No usage restrictions beyond attribution.

质量与使用场景

评分由模型作者或独立评测方发布——衡量质量而非吞吐量,并非我们实测。

最适合software engineeringagentic codingrepo leveltest driven dev
基准分数来源
SWE-bench Verified71.6 / 100 %已发布
HumanEval91.2 / 100 %已发布
LiveCodeBench v667.8 / 100 %已发布

我的 GPU 能运行 Devstral-2 123B 吗?

Devstral 的其他尺寸

Devstral-2 123B — 常见问题

How much VRAM does Devstral-2 123B need?

About 75 GB at Q4_K_M — quantized weights plus framework overhead, before any KV cache. The cache grows with context length and is added on top; the table above folds it in. Apple Silicon counts unified memory toward the same figure.

Does Devstral-2 123B run on an RTX 4090 (24 GB)?

No. Devstral-2 123B needs about 75 GB at Q4_K_M, more than a single RTX 4090's 24 GB. It needs a larger card, several GPUs, or Apple Silicon with enough unified memory — or it runs with part of the weights offloaded to system RAM, which is much slower.

How do I run Devstral-2 123B locally?

Install Ollama and run `ollama run devstral:123b`. That pulls the weights and starts a local OpenAI-compatible endpoint; after the download nothing leaves the machine.

What other sizes does Devstral come in?

Devstral-2 123B (75 GB), Devstral 2 22B (14 GB), Devstral Small 2505 24B (15 GB), Devstral Small 2 24B (15 GB). Every size shares the family's training and licence; the larger ones score higher and need proportionally more memory.