Mistral released Leanstral 1.5 on July 2, a free and open-source model that automates formal verification using Lean 4, giving developers and mathematicians a practical way to prove software correctness and solve advanced theorems at a fraction of the cost of existing tools.
The model uses a mixture-of-experts architecture with 119 billion total parameters and just 6 billion active at inference time, keeping it efficient enough to run on standard hardware. Mistral trained it across three stages — mid-training, supervised fine-tuning, and reinforcement learning with CISPO — in two environments: a multi-turn theorem-proving sandbox and a code agent environment that simulates real filesystem operations.
Benchmark results
Leanstral 1.5 is the first publicly available model to fully saturate miniF2F, a standard theorem-proving benchmark, achieving 100% on both validation and test sets. On PutnamBench — drawn from the William Lowell Putnam Mathematical Competition, one of the most demanding undergraduate math contests — it solves 587 of 672 problems. It also reaches state-of-the-art scores of 87% on FATE-H and 34% on FATE-X.
The cost gap is notable: Mistral reports approximately $4 per problem solved, compared to $300 or more for comparable systems.
Real-world impact
Beyond benchmarks, Mistral ran Leanstral 1.5 against 57 open-source repositories and uncovered five previously unknown bugs, including a critical integer overflow in a Rust library’s zigzag decoding function. In a longer task, the model proved O(log n) time complexity for AVL tree operations after working through 2.7 million tokens of proof state.
Access
Weights are available on Hugging Face under the Apache-2.0 license. A free API endpoint (leanstral-1-5) is accessible through Mistral’s platform, with integration into Mistral Vibe. According to Mistral, the free endpoint is scheduled to close on September 30, 2026.