Introduction to Leanstral 1.5
Leanstral 1.5, a free Apache-2.0 licensed model with 119B total and only 6B active parameters, delivers a performance upgrade that makes formal verification more powerful and accessible than ever [1]. It saturates miniF2F, solves 587/672 PutnamBench problems, and achieves a new state-of-the-art of %87 on FATE-H and 34% on FATE-X [2].
Training Leanstral 1.5
Leanstral 1.5 goes through a three-stage process: mid-training, supervised fine-tuning, and reinforcement learning with CISPO [3]. It leverages extensive training on two RL environments: the multiturn environment and the code agent environment.
Evaluation
We evaluate Leanstral on the following benchmarks: miniF2F, PutnamBench, FATE-H, and FATE-X [4]. It shows the strongest test-time scaling we have seen from a formal-reasoning model.
Code Verification Case Studies
While being primarily trained for mathematics, Leanstral 1.5 exhibits strong abilities in code verification. We present 2 critical case studies to demonstrate its impact: AVL Trees: Proving Time Complexity and Bug Discovery: Finding Hidden Flaws [5].
Get Started
Leanstral 1.5 has a Apache-2.0 license. The weights can be found on Huggingface, while also being available now as a free API endpoint as leanstral-1-5. We recommend using it in Mistral Vibe.
Sources
- Leanstral 1.5: Proof Abundance for All. Mistral AI.
- Leanstral 1.5: Proof Abundance for All. Hacker News.
- Sophia Yang, Ph.D. on X: Learn more about how it's trained:
- Mistral ships Leanstral 1.5 for Lean 4 proof work, free in Labs. AI Weekly.
- Leanstral 1.5: Proof Abundance for All. Mistral AI.
