Mistral AI has released Leanstral 1.5, an open-source model under the Apache 2.0 licence designed for formal verification within the Lean 4 programming language. This tool targets the rigorous validation of mathematical proofs and software correctness. The model achieved a perfect score on the miniF2F benchmark, which covers problems ranging from high school level to math olympiad difficulty. It also solved 587 out of 672 problems on the PutnamBench competition. On the algebra benchmarks FATE-H and FATE-X, which test master’s and doctoral-level tasks in group theory and ring theory, it secured top results of 87 and 34 percent respectively.
Beyond mathematics, the model demonstrated practical utility in code verification during a hands-on test. It scanned 57 open-source repositories and identified five previously unknown bugs, including an overflow issue in the Rust library varinteger. The system is available through Hugging Face and a free API, following training that involved mid-training, supervised fine-tuning, and reinforcement learning.
* Released under Apache 2.0 licence for unrestricted commercial use
* Scores 587 of 672 problems on PutnamBench
* Detected an overflow bug in the Rust library varinteger
