Mistral releases Leanstral: Lean4's first open-source code agent that generates code and outputs formal proofs simultaneously.

This article is machine translated
Show original

According to 1M AI News , Mistral AI today released Leanstral, the first open-source code agent specifically designed for the formal verification tool Lean 4. The core bottleneck in AI code generation is human review; Leanstral bypasses this step by generating code and simultaneously outputting formal proofs that can be automatically verified by Lean 4. The model employs a sparse MoE architecture with 120B total parameters and 6B activation parameters, is open-sourced under Apache 2.0, and features specific training optimizations for lean-lsp-mcp. It can be started with zero configuration in the Mistral Vibe (command `/leanstall`) or called via the free API endpoint `labs-leanstral-2603`, supporting self-deployment of downloaded weights.

Mistral also released a new evaluation benchmark, FLTEval, using the Fermat's Last Theorem formalization project from the Lean 4 community as its testbed. Cost comparison: Leanstral pass@2 scored 26.3 at $36, surpassing Claude Sonnet 4.6 (23.7 points) at $549; pass@16 scored 31.9 at $290, leading Sonnet by 8 points, while Claude Opus 4.6 requires $1,650 to reach 39.6. Among open-source models, Qwen3.5-397B-A17B requires 4 runs to reach 25.4 points, still lower than Leanstral pass@2.

Source
Disclaimer: The content above is only the author's opinion which does not represent any position of Followin, and is not intended as, and shall not be understood or construed as, investment advice from Followin.
Like
Add to Favorites
Comments