Mistral AI has released Leanstral 1.5, a code agent model built for Lean 4, the programming language used to write and verify mathematical proofs. Unlike many frontier AI releases, the model is free and open under an Apache-2.0 license, meaning developers can use, modify, and deploy it without licensing fees.
According to MarkTechPost, Leanstral 1.5 posts strong results on two well-known proof benchmarks. It "saturates" miniF2F — effectively topping out that test set — and solves 587 of 672 problems on PutnamBench, a collection derived from the notoriously difficult Putnam mathematics competition.
The model uses a mixture-of-experts design, a technique where only part of the network activates for any given input. MarkTechPost reports it totals 119 billion parameters but activates just 6.5 billion per token, a setup that aims to deliver large-model capability at a lower computing cost per query. MarkTechPost says its coverage also breaks down the architecture, benchmarks, real bug-finding case studies, and deployment.
The release drew attention on Hacker News, where a post titled "Leanstral 1.5: Proof Abundance for All" reached the site's front page with 105 points and 30 comments.
Why it matters: tools that can automatically generate and check formal proofs could make rigorous mathematics and verified software far more accessible, and releasing such a capable model under a permissive open license puts that ability directly in the hands of researchers, students, and developers.