Leanstral 1.5 is a free Apache-2.0 licensed model with 6B active parameters that delivers significant upgrades to formal verification capabilities. The system saturates mini F2F completely and solves 587 out of 672 Putnam Bench problems, establishing new state-of-the-art results with 87% on FATE-H and 34% on FATE-X benchmarks.

Training follows a three-stage process involving mid-training, supervised fine-tuning, and reinforcement learning with CISPO. Unlike competitors that cost $300 or more per problem, Leanstral achieves results at roughly $4 per problem while demonstrating strong test-time scaling that continues improving with increased token budgets.

The model excels in practical applications, proving AVL tree time complexity guarantees over 2.7 million tokens and discovering 5 previously unknown bugs across 57 tested repositories. It operates as a code agent that edits files, runs bash commands, and uses the Lean language server for real-time feedback on goals and errors.

Fully open-sourced on Hugging Face with a free API endpoint, Leanstral 1.5 makes rigorous formal methods accessible for real-world proof engineering in Lean 4, showing that automated verification can already catch edge cases missed by traditional testing approaches.