Mistral ships Codestral 2 - Apache 2.0 licensed code model with 380K first-week downloads

Mistral released Codestral 2 in April, a 22B parameter model built specifically for code generation and editing. The model attracted 380,000 downloads in its first week, making it one of the fastest-adopted code models of the month.

The Apache 2.0 license is the headline feature. Unlike many code models that come with custom licenses containing revenue caps, usage restrictions, or registration requirements, Apache 2.0 has none of that. You can use Codestral 2 in commercial products, modify it, redistribute it, and build services on top of it with no restrictions whatsoever. There is no small print, no revenue threshold, and no requirement to notify Mistral about your use.

At 22B parameters, the model is practical for local deployment. A single RTX 3090 or 4090 handles it comfortably with quantization. Inference is fast enough for interactive use, which matters for coding tools where latency directly affects the developer experience. Autocomplete suggestions appear in under a second on modern hardware.

Mistral positioned Codestral 2 as a direct competitor to proprietary code models. The benchmarks show it competing with models from larger labs on standard code generation tasks, including HumanEval, MBPP, and real-world repository-level coding tests. Real-world performance always depends on your specific use case and codebase, so testing on your own projects is more informative than benchmark scores.

The model supports all major inference engines. Ollama, vLLM, llama.cpp, LM Studio, and text-generation-webui all work without configuration changes. If you already have a local inference setup, adding Codestral 2 is as simple as pulling the model weights with ollama pull codestral:22b.

For anyone running AI coding agents, Codestral 2 fills a specific gap: a capable code model with zero licensing risk. If a client or employer asks what license your AI coding tool uses, "Apache 2.0" is the simplest possible answer. No lawyers needed, no compliance review required.

The release is part of a broader trend where specialized models outperform general-purpose ones on targeted tasks. A 22B code model trained specifically on code can match or beat a 70B general model on programming benchmarks, while being much cheaper and faster to run locally.