Introducing Mistral Code

Mistral AI wants to solve a problem that's been quietly stalling AI coding tools inside big companies: nobody trusts the black box enough to let it near proprietary code. Mistral Code, launched today in private beta for JetBrains and VSCode users, is the company's answer — a coding assistant where the models, the plugin, and the enterprise controls all come from a single vendor, under a single SLA, with every line of code staying inside the customer's own infrastructure.

The product runs on four Mistral models working in tandem: Codestral handles autocomplete, Codestral Embed does code search, Devstral tackles agentic multi-step tasks, and Mistral Medium powers the chat interface. That's a deliberate split from the usual copilot setup, where a company might license a plugin from one vendor, a model from another, and host infrastructure with a third — a arrangement that Mistral says routinely dies in the details once CISOs and platform leads start asking hard questions. Mistral spoke with VPs of engineering and security chiefs and found the same four complaints kept surfacing: copilots that can't touch internal repos, models nobody can customize, tools that stop at suggesting code rather than finishing tickets, and support spread thin across vendors who all point fingers at each other.

What's notable is how far Mistral is pushing the