Codestral

Mistral AI has entered the coding-model race with Codestral, a 22-billion-parameter open-weight system trained specifically to read, write, and complete code. Unlike the company's general-purpose Mistral models, this one is built around a single job: helping developers ship code faster, in whatever language they happen to be using. And the language list is long — more than 80 of them, from the obvious picks like Python, Java, and JavaScript down to niche entries like Swift and Fortran.

What sets Codestral apart isn't just breadth, it's reach. The model handles a 32,000-token context window, dwarfing the 4k-to-16k windows typical of competing code models. On RepoBench, a benchmark designed to test long-range code completion across entire repositories rather than isolated snippets, that extra context translates into a real performance edge. Mistral also ran it through HumanEval, MBPP, CruxEval, and the Spider SQL benchmark, and reports that Codestral holds its own against pricier, heavier models — including on fill-in-the-middle tasks, where it was measured directly against DeepSeek Coder 33B.

The rollout comes with two ways to actually use the thing. There's a new dedicated endpoint, codestral.mistral.ai, aimed at people building IDE plugins — it's free for eight weeks, though gated behind a waitlist to keep quality steady under load. Then there's the standard api.mistral.ai route, billed per token, meant for research, batch jobs, or third-party apps that don't expect users to bring their own keys. Mistral is also offering self-deployment for companies that want the model running on their own infrastructure.

The integration list reads like a who's-who of developer tooling: LlamaIndex and LangChain for building agentic apps, plus Continue.dev and Tabnine for in-editor autocomplete and chat inside VSCode and JetBrains. JetBrains itself ran Codestral against its Kotlin-HumanEval benchmark and says it scored 73.75 at T=0.2, edging out GPT-4-Turbo's 72.05 and clearing GPT-3.5-Turbo by nearly 20 points. Sourcegraph's team, meanwhile, is more focused on latency than raw accuracy — for autocomplete tools like Cody, shaving milliseconds off response time matters more than squeezing out another percentage point of benchmark score.

Licensing is the one place Mistral hasn't gone fully open. Codestral sits under the company's new Non-Production License, meaning it's free to download from Hugging Face for research and testing, but any commercial use requires reaching out for a paid license. That's a familiar pattern for Mistral: open enough to build goodwill and get feedback from the developer community, closed enough to keep a revenue lever for actual production use.