majiayu000 Mistral
Mistral AI focuses on efficiency and coding capabilities. Their "Mixture of Experts" (MoE) architecture (Mixtral) changed the game.
When to Use
- 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).
Core Concepts
MoE (Mixture of Experts)
Only a subset of parameters (experts) are active per token. High quality, low compute.
Codestral
A model trained specifically on 80+ programming languages.
Le Chat
Mistral's chat interface (chat.mistral.ai).
Best Practices (2025)
Do:
- Use
codestral-mamba: For infinite context window coding tasks (linear time complexity). - 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.