Leanstral: Open Foundation for Trusted Programming
Mistral
Anthropic
Alibaba/Qwen
Moonshot AI
Mistral AI has unveiled Leanstral, the first open-source agent for the Lean 4 proof assistant. The model, with 6 billion active parameters, tackles formal verification tasks, outperforming larger open models in efficiency and achieving competitive results at a significantly lower cost compared to Anthropic's Claude family.
Mistral AI has announced Leanstral, the first open-source agent for Lean 4, a proof assistant capable of handling complex mathematical objects and software specifications. The model uses a sparse architecture (6 billion active parameters) and is optimized for formal proof tasks, supporting parallel inference with Lean as a verifier. Leanstral is available under the Apache 2.0 license, in the Mistral Vibe environment, and through a free API. To evaluate realistic scenarios, the company developed a new benchmark, FLTEval. In tests, Leanstral scored 26.3 with pass@2, surpassing Claude Sonnet (23.7) at a cost of $36 versus $549. With pass@16, it achieved 31.9 points, second only to Claude Opus 4.6 (39.6), which costs 92 times more ($1650). Among open models, Leanstral outperformed Qwen3.5-397B-A17B (25.4 points over 4 passes), GLM5-744B-A40B (16.6), and Kimi-K2.5-1T-32B (20.1) in just a single pass. In case studies, the model successfully converted code from Rocq to Lean and solved real-world problems from Stack Exchange related to Lean version updates.
Source: Mistral AI —
original