Mistral AI Releases Leanstral 1.5, an Open-Source Model That Catches Real Bugs in Code

Mistral AI released Leanstral 1.5 on July 4, 2026, an open-source model built for formal verification in the Lean 4 programming language. The model is available under the Apache 2.0 license and can be accessed via Hugging Face or a free API.

On the miniF2F benchmark, which tests formal math problems from high school to olympiad difficulty, Leanstral 1.5 scores 100 percent. It solves 587 of 672 problems on PutnamBench, a collection of problems from the Putnam mathematics competition. On the algebra benchmarks FATE-H and FATE-X, which cover master's and doctoral-level tasks in group theory and ring theory, it achieves top scores of 87 percent and 34 percent respectively.

Real-world bug hunting across 57 repositories

While the model was trained primarily for mathematics, Mistral says it also performs well at code verification. In a hands-on test, Leanstral 1.5 scanned 57 open-source repositories and identified five previously unknown bugs. One of the bugs was an overflow vulnerability in the Rust library varinteger. The model's training involved mid-training, supervised fine-tuning, and reinforcement learning.

• Leanstral 1.5 tops the open-source field on PutnamBench, FATE-H, and FATE-X benchmarks.
• Only the closed-source Aleph Prover outperforms it on PutnamBench.
• The model is designed for Lean 4, a language used to formally verify mathematical proofs and software correctness.
• It is available under the permissive Apache 2.0 license.

Implications for software verification and open-source security

The ability to catch real bugs in production code marks a shift from theoretical benchmarks to practical utility. Formal verification has historically been a niche skill, but models like Leanstral 1.5 lower the barrier for developers to check their code for correctness. Mistral's decision to release the model as open-source means that any team can integrate it into their development pipeline without licensing costs.

What comes next is likely broader adoption of AI-assisted formal verification in open-source projects. If Leanstral 1.5 can find five bugs in 57 repositories, scaling that scan to thousands of projects could uncover many more. Mistral has not announced a timeline for a larger model, but the company continues to invest in the intersection of AI and formal methods.