Mistral's Leanstral 1.5 is an open model that proves theorems and finds real bugs

Mistral released Leanstral 1.5, an open-weight model built for formal proofs in Lean 4, and the numbers are unusually concrete. It reaches 100% on the miniF2F validation and test sets and solves 587 of 672 problems on PutnamBench, a benchmark drawn from competition mathematics. The model has 119B total parameters but activates only 6B at a time, and the weights sit on Hugging Face under an Apache-2.0 license, with a free API endpoint.

What makes it worth attention beyond the benchmark table is that it works as an agent. It edits proof files and reads feedback from the Lean language server to refine its attempts, and it scales with compute: given a budget of 4 million tokens instead of 50,000, it goes from solving 44 problems to 587. On one AVL tree proof it worked through more than 2.7 million tokens across 22 proof compactions. It also found a real integer overflow bug in datrs/varinteger, a live Rust repository, while verifying code rather than passing exercises. Mistral says that on FLTEval it beats Claude Opus 4.6 at roughly one-seventh the cost.

Why it matters

If you write software where correctness has to be proven rather than only tested, an open model that runs formal verification agentically, and finds real bugs while doing it, is something you can now try without paying a frontier lab. The cost gap against a closed model on the same task is the figure to weigh against your own workload before committing.