Leanstral: open-source foundation for trusted programming
Mistral
Anthropic
Alibaba/Qwen
Moonshot AI
Mistral AI has released Leanstral, the first open-source code agent for Lean 4, a proof assistant. Leanstral is efficient with 6B active parameters and outperforms larger open-source models and some Claude models at a fraction of the cost, aiming to verify code generation.
Mistral AI released Leanstral, the first open-source code agent designed for Lean 4, a proof assistant capable of expressing complex mathematical objects and software specifications. Leanstral uses a highly sparse architecture with 6B active parameters, leverages parallel inference with Lean as a perfect verifier, and supports arbitrary MCPs. It is released under Apache 2.0 license, available in Mistral Vibe, via a free API endpoint, and as downloadable weights. Evaluated on FLTEval, a new evaluation suite for realistic proof engineering, Leanstral-120B-A6B achieved a score of 26.3 at pass@2, beating Sonnet (23.7) while costing $36 vs. $549, and reaching 31.9 at pass@16. It also outperformed large open-source models like GLM5-744B-A40B and Kimi-K2.5-1T-32B. Case studies include successfully answering a Stack Exchange question about breaking changes in Lean 4.29.0-rc6 and converting Rocq program definitions to Lean and proving properties.
- Сокращения
- MCP = Model Context Protocol — Протокол контекста модели
- API = Application Programming Interface — Программный интерфейс приложения
- PR = Pull Request — Запрос на слияние
- OSS = Open-Source Software — Программное обеспечение с открытым исходным кодом
Source: Mistral AI —
original