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.