Announcing Codestral 25.08 and the Complete Mistral Coding Stack for Enterprise
Mistral AI
Mistral just bundled its coding models into one enterprise stack: Codestral, Devstral, embeddings, and an IDE plugin, all self-hostable. Banks and regulated firms can finally run AI coding tools without sending code to someone else's cloud.
Based on reporting by Mistral AI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Enterprise coding assistants have a dirty secret: most of them only work if you're willing to send your codebase to a third-party SaaS server. For a bank in Spain or a railway operator in France, that's a non-starter. Mistral's answer, announced this week, is to stop selling point tools and start selling a full stack that a security team can actually approve.
The centerpiece update is Codestral 25.08, the latest version of Mistral's fill-in-the-middle completion model. The company claims a 30% jump in accepted completions, 10% more retained code after suggestions, and half as many runaway generations compared to the prior release — numbers pulled from live usage in production codebases, not just academic benchmarks. Sitting alongside it is Codestral Embed, a code-specific embedding model that Mistral says beats OpenAI's and Cohere's offerings on retrieval tasks, even when compressed down to 256 dimensions or INT8 precision to save storage.
Then there's Devstral, the agentic layer built on the OpenHands scaffold, which is where things get more ambitious than autocomplete. Devstral Small, a 24B open-weight model under Apache 2.0, runs on a single RTX 4090 and scores 53.6% on SWE-Bench Verified. Devstral Medium, available through Mistral's API and enterprise deals, hits 61.6% — beating Claude 3.5 and GPT-4.1-mini by Mistral's own measurements. The pitch is that a developer can ask it to swap out a deprecated retry function across three services, and it'll do the cross-file edits, write the changelog, and draft the pull request for a human to review.
All of this gets wired into JetBrains and VS Code through a plugin called Mistral Code, which is the part enterprises will actually care about most. It supports on-prem deployment (general availability is slated for Q3), skips mandatory telemetry, and gives platform teams SSO, audit logs, and usage dashboards through the Mistral Console. That's the boring infrastructure work that determines whether a CISO signs off, and it's clearly where Mistral is putting its differentiation rather than just chasing benchmark scores.
Customers cited include Capgemini, which is running the stack across delivery teams in defense and energy, the Spanish bank Abanca doing a fully self-hosted deployment for data residency reasons, and SNCF using agents to chip away at legacy Java systems with a human still in the loop. None of that is flashy, but it's exactly the kind of unglamorous adoption that determines whether AI coding tools become infrastructure or stay a demo.
My take — AI-written commentary, not fact-checked reporting
This is the correct instinct: enterprises don't want a smarter autocomplete, they want something their compliance team won't kill in week two, and Mistral is betting the on-prem, open-weight angle is worth more than chasing another benchmark point past Claude or GPT. It also doubles as a quiet argument for European tech sovereignty — a France-based company selling self-hosted AI to a French railway and a Spanish bank isn't subtle, and I think that's smart positioning rather than mere flag-waving.
Read more about this at: Mistral AI