Mistral releases Leanstral 1.5, an Apache-2.0 Lean 4 code agent solving 587/672 PutnamBench problems

Mistral released Leanstral 1.5, an Apache-2.0-licensed code agent model specialized for Lean 4 and automated theorem proving, released under the banner 'Proof Abundance for All.' Part of the Mistral Small 4 family, it reports solving 587 of 672 PutnamBench problems and achieving state-of-the-art results on the FATE-H (87%) and FATE-X (34%) algebra benchmarks — strong numbers for a small, permissively licensed model aimed at formal mathematics.

Distribution is a core part of the story: the model is fully open-sourced on Hugging Face and available through a free API endpoint, reinforcing Mistral's open-weight positioning and Hugging Face's role in democratizing frontier-adjacent research models. Automated theorem proving is a niche but high-value domain — it's a proving ground for reliable, verifiable AI reasoning that can't hallucinate its way past a formal checker.

The release complements Mistral's broader week (its 'Europe's OpenAI rival' narrative and rumored $3.5B raise) and signals the company competing on research credibility, not just enterprise deployment. Competitively, it lands against DeepSeek's open-weight V4 and other math/reasoning specialists, and continues the open-vs-closed tension that dominated developer chatter this week.

Caveats: PutnamBench and FATE are specialized benchmarks, and theorem-proving skill doesn't necessarily transfer to general coding. What to watch: independent verification of the PutnamBench numbers, community adoption via the free API, and whether Leanstral feeds into Mistral's forthcoming general open-weight model.