# Mistral Medium 3.5：定位胜于基准测试

- 来源：Chubby♨️ (@kimmonismus)
- 发布时间：2026-04-30 01:43
- AIHOT 分数：51
- AIHOT 链接：https://aihot.virxact.com/items/cmokdd64m00qtsl4f85ng9qoi
- 原文链接：https://x.com/kimmonismus/status/2049545016784413005

## AI 摘要

Mistral Medium 3.5是MistralAI的新旗舰模型，以公共预览版发布。它整合指令遵循、推理和编码能力，采用128B密集参数和256k上下文窗口，支持可配置推理努力。模型定位比基准测试更关键，比较对象包括Kimi、Qwen、GLM和Claude Sonnet，而非GPT或Gemini。随着Aleph Alpha被Cohere收购，Mistral成为唯一非美国、非中国的尖端实验室，以开源权重和修改的MIT许可证发布。模型在推理效率与一致性间权衡，Collie分数达95.8领先，目标不是原始推理，而是成为生产中可靠遵循指令的模型，体现欧洲企业定位。它是Mistral Vibe和Le Chat的新默认模型。

## 正文

Mistral Medium 3.5 is interesting less for the benchmarks and more for the positioning. Look at who they're comparing against： Kimi， Qwen， GLM， Claude （Sonnet）. Not GPT， not Gemini. And i dont mean that in a negative way！

With Aleph Alpha being acquired by Cohere last week， Mistral is now the only non-US， non-Chinese lab still in the frontier conversation. At 128B dense with open weights， they're making a different bet than the Chinese MoE models in that chart （which activate only 17-40B params despite being 400B-1T total）.

Mistral is trading inference efficiency for consistency. The Collie score （95.8， best in class by a wide margin） tells you where they're aiming： not raw reasoning， but the most reliable model to actually follow instructions in production. That's a European enterprise pitch， not a benchmark race.

Very solid release from Mistral！

### 引用推文

> Mistral Vibe：Mistral Medium 3.5, a new flagship model in public preview by @MistralAI that merges instruction-following, reasoning, and coding into a single 128B dense model...