mistral

name	mistral
description	Mistral AI efficient open models. Use for efficient AI.

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.

References

• Mistral AI Documentation