Mistral released Codestral 2, their second model built specifically for writing and understanding code. At 22 billion parameters, it's large enough to be highly capable but compact enough to run on reasonable hardware. The most useful feature for anyone working with large codebases is the 256,000-token context window: you can feed it an entire project before asking it to do anything. It supports function calling (so it can act as an agent that calls tools) and JSON mode for structured outputs. The model weights are available to download, and from day one it works on Ollama, making it trivial to install locally on any Mac or PC with enough RAM. Compared to closed models like GPT-4o, Codestral 2 is the top choice for anyone who wants power without depending on a cloud subscription.
Companies
Mistral AI
Tools
Codestral 2, Ollama
Tags
Sources