While everyone piles into general-purpose coding agents, Mistral went the other way. On June 30 it shipped Leanstral 1.5, a model built for one narrow, hard thing: automated theorem proving and autoformalization in Lean 4. It replaces the March original, which is now retired.
What it actually does
Leanstral 1.5 is a 119B-parameter MoE model with only 6.5B active, and a 256k context window big enough to swallow a proof file plus its surrounding code. Its job is translating human math and specs into Lean 4 the compiler can check. No hand-waving: the proof either compiles or it doesn’t. That binary verifier is the whole point — this is the foundation for trustworthy “vibe coding” and machine-checked math, not another autocomplete.
Why it’s worth a look
It’s free right now in Mistral Labs, usable straight from the console playground with the full 256k context — good for auto-formalizing statements, closing proof goals, or verifying generated code. The March version already beat Claude Sonnet by 8 points at pass@16 at 15x lower cost. Niche, sure. But correctness you can mechanically prove is a bet almost nobody else is making.
You Might Also Like
- Cloudrouter Gives Your ai Coding Agent its own Cloud Machine and Thats a big Deal
- Vet by Imbue the Safety net Your ai Coding Agent Desperately Needs
- P0 Bepurple ai Finally an ai Coding Agent That Actually Ships Whole Features
- Mistral Small 4 Packs 119b Parameters Into 6b Active and it Does Everything
- Google ai Studio 2 0 Ships Full Stack Vibe Coding With Firebase and its Free
Top AI Product 
Leave a comment