French AI startup Mistral AI has unveiled Leanstral, an innovative AI model designed to aid in mathematical proofs and software specification verification. Released on March 17, 2026, Leanstral operates as an open-source AI agent compatible with the formal proof tool Lean 4, aiming to enhance ‘proof engineering’—a discipline focused on rigorously ensuring the correctness of mathematical computations and programming.
While artificial intelligence has made significant strides in reasoning, mathematical proof generation, and coding, the final checks and validations still require human oversight to ensure accuracy. As mathematical research and software complexity increase, the manual verification process becomes a bottleneck, often hindering engineering efficiency. Mistral AI’s vision for Leanstral is to create coding agents that not only perform tasks but also formally validate the correctness of their implementations, thereby streamlining the verification process.
Leanstral is distinguished as the first open-source code agent specifically designed for Lean 4, which is widely utilized in mathematical research and software verification fields. The model employs a Mixture-of-Experts (MoE) architecture optimized for proof engineering tasks. This innovative design allows Leanstral to selectively leverage specialized modules, which improves performance while keeping computational costs low by utilizing only a fraction of its total parameters during calculations. By integrating Lean as a verifier while generating and validating multiple inference outcomes, Leanstral demonstrates superior performance and cost efficiency compared to existing closed-source competitors.
In performance benchmarks, Leanstral has outperformed major open-source models in formal proof completion and correct mathematical concept definitions, as measured by the newly introduced FLTEval score. For instance, the model achieved a score of ‘26.3’ in two attempts, surpassing the leading open model, Qwen3.5 397B-A17B, which scored ‘25.4’ in four attempts. Leanstral further improved its score to ‘29.3’ after four attempts, marking a significant advancement in the realm of AI-assisted proof verification.
When it comes to cost-effectiveness, Leanstral stands out in comparison to other coding agents such as Claude Opus 4.6, Sonnet 4.6, and Haiku 4.5. While Claude Opus boasts a high score of ‘39.6’, its execution costs are approximately 92 times greater than those incurred by Leanstral for equivalent performance. This highlights Leanstral’s potential to deliver high-quality results at a fraction of the cost, making it an appealing option for researchers and developers.
Leanstral has been released under the Apache 2.0 license through its agent mode within Mistral Vibe and is also accessible via a free API endpoint. This openness allows researchers and developers to freely use and modify the tool. In addition, a technical report detailing the training methodology for Leanstral and a new evaluation suite, FLTEval, are anticipated for release. These developments signal Mistral AI’s commitment to advancing the utility of AI in formal verification and proof engineering.
See also
Amazon and NVIDIA Collaborate to Launch AI-Powered In-Car Assistants for Automakers
Anthropic Hires Safety Manager to Combat AI Risks from Chemical and Explosive Threats
Anthropic Hires Explosives Expert to Enhance AI Safety Against Weapon Misuse
Musk’s Grok Feature Paywalled for Free Users Amid xAI’s Struggles and Leadership Changes
OpenAI Secures Amazon AWS Deal for Classified Work, Leaving Anthropic Behind
