AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Leanstral: Open-Source foundation for trustworthy vibe-coding

Collected Oct 1, 2026

Mistral AI announced Leanstral, which it describes as the first open-source code agent designed for Lean 4, a proof assistant capable of expressing complex mathematical objects and software specifications such as properties of Rust fragments. The company said Leanstral has 6B active parameters and was trained for operating in realistic formal repositories, rather than acting as a wrapper around generalist models or focusing on single math problems.

Mistral released Leanstral weights under an Apache 2.0 license, made it available in an agent mode within Mistral vibe, and offered it through a free API endpoint. It said it will also release a tech report detailing its training approach and a new evaluation suite, FLTEval, to move evaluations beyond competition math. Leanstral supports arbitrary MCPs through vibe and was specifically trained to achieve maximal performance with lean-lsp-mcp, according to Mistral.

For evaluation, Mistral benchmarked Leanstral on completing all formal proofs and correctly defining new mathematical concepts in each pull request to the FLT project, comparing it against leading coding agents (Claude Opus 4.6, Sonnet 4.6, Haiku 4.5) and open-source models (Qwen3.5 397B-A17B, Kimi-K2.5 1T-A32B, GLM5 744B-A40B).

Mistral said Leanstral-120B-A6B has a significant efficiency advantage over much larger open-source peers. It said GLM5-744B-A40B and Kimi-K2.5-1T-32B capped their FLTEval scores at approximately 16.6 and 20.1 respectively, while Leanstral outperformed both with a single pass. It said Qwen3.5-397B-A17B, the strongest open-source competitor shown, required 4 passes to reach 25.4, while Leanstral reached 26.3 at pass@2 and 29.3 at the same cost level.

Against the Claude family, Mistral said Leanstral pass@2 scored 26.3, beating Sonnet by 2.6 points while costing $36 to run versus Sonnet's $549, and that at pass@16 Leanstral reached 31.9, beating Sonnet by 8 points. It said Claude Opus 4.6 remains the leader in quality but carries a cost of $1,650, which Mistral described as 92 times higher than running Leanstral. Mistral said it used Mistral Vibe as the scaffold with no modifications specifically for the evaluation.

Leanstral is available today, Mistral said, through Mistral Vibe using the /leanstall command or vibe --agent lean, via the labs-leanstral-2603 API endpoint, and as downloadable Apache 2.0 licensed weights.

Read at Mistral AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt