Mistral's Leanstral 1.5 Crushes Formal Verification Benchmarks, Finds Real Bugs

Mistral has released Leanstral 1.5, an open-source AI model that saturates the miniF2F benchmark and achieves new bests on advanced math challenges like PutnamBench and FATE-H/X. The model packs 119B total parameters but runs with only 6B active, delivering elite performance at a fraction of the cost of competitors.

For builders and founders, this makes formal verification—a method to mathematically prove code correctness—practical and accessible. Leanstral 1.5 can verify complex properties and automatically detect bugs in real codebases, as shown by its discovery of unreported flaws in open-source repositories.

Trained through a three-stage process including reinforcement learning with CISPO, the model exhibits strong test-time scaling: on PutnamBench, it solved 587 out of 672 problems at a 4-million-token budget, with performance climbing steadily as more compute is applied.

In case studies, it proved the O(log n) time complexity of AVL trees over a 2.7-million-token session and built a pipeline that found genuine bugs in 11 out of 57 tested code repositories, including an integer overflow in a Rust library that traditional testing would likely miss.

Available under an Apache-2.0 license on Hugging Face and as a free API endpoint, Leanstral 1.5 lowers the barrier to entry for formal methods, letting developers tackle theorems, debug proofs, or harden code with minimal cost.