NEWS · MODELS · #464
Leanstral 1.5: open-source 6B-active-parameter model advances formal verification
Leanstral 1.5 is an Apache-2.0 open-source model (119B total, 6B active parameters) released for Lean 4 proof engineering; it uses mid-training, supervised fine-tuning, and reinforcement learning with CISPO and is available via Hugging Face and a free API. The model saturates miniF2F, solves 587/672 PutnamBench problems, achieves 87% on FATE-H and 34% on FATE-X, scales strongly with token budget, improves FLTEval pass rates, and found 5 previously unknown bugs across 57 tested repositories.
KEY POINTS
- Leanstral 1.5 is an Apache-2.0 open-source model (119B total, 6B active parameters) released for Lean 4 proof engineering; it uses mid-training, supervised fine-tuning, and reinforcement learning with CISPO and is available via Hugging Face and a free API.
- The model saturates miniF2F, solves 587/672 PutnamBench problems, achieves 87% on FATE-H and 34% on FATE-X, scales strongly with token budget, improves FLTEval pass rates, and found 5 previously unknown bugs across 57 tested repositories.
- This release delivers state-of-the-art formal-reasoning performance in an openly available, cost-efficient model, making practical formal verification and automated proof engineering more accessible.
WHY IT MATTERS
This release delivers state-of-the-art formal-reasoning performance in an openly available, cost-efficient model, making practical formal verification and automated proof engineering more accessible.