mistral
Compare original and translation side by side
🇺🇸
Original
English🇨🇳
Translation
ChineseMistral
Mistral
Mistral AI focuses on efficiency and coding capabilities. Their "Mixture of Experts" (MoE) architecture (Mixtral) changed the game.
Mistral AI 专注于效率与编码能力。其“混合专家”(MoE)架构(Mixtral)颠覆了行业格局。
When to Use
适用场景
- Coding: Mistral Large 2 (Codestral) is specifically optimized for code generation.
- Efficiency: Mixtral 8x7B offers GPT-3.5+ performance at a fraction of the inference cost.
- Open Weights: Apache 2.0 licenses (for smaller models).
- 编码任务:Mistral Large 2(Codestral)专为代码生成优化。
- 高效推理:Mixtral 8x7B 能达到GPT-3.5+的性能水平,而推理成本仅为其一小部分。
- 开源权重:小模型采用Apache 2.0许可证。
Core Concepts
核心概念
MoE (Mixture of Experts)
MoE(混合专家)
Only a subset of parameters (experts) are active per token. High quality, low compute.
每个token仅激活部分参数(专家模块),兼顾高质量与低算力消耗。
Codestral
Codestral
A model trained specifically on 80+ programming languages.
一款针对80余种编程语言训练的模型。
Le Chat
Le Chat
Mistral's chat interface ().
chat.mistral.aiMistral的聊天界面()。
chat.mistral.aiBest Practices (2025)
2025年最佳实践
Do:
- Use : For infinite context window coding tasks (linear time complexity).
codestral-mamba - Deploy via vLLM: Mistral models run exceptionally well on vLLM.
Don't:
- Don't ignore small models: Mistral NeMo (12B) is surprisingly capable for RAG.
推荐做法:
- 使用:适用于无限上下文窗口的编码任务(线性时间复杂度)。
codestral-mamba - 通过vLLM部署:Mistral模型在vLLM上的运行表现极为出色。
不推荐做法:
- 不要忽视小模型:Mistral NeMo(12B)在RAG场景中的表现出人意料地出色。