Leanstral 1.5がminiF2Fで100%を達成——さらに実際のバグも検出
MistralがLeanstral 1.5をリリース。miniF2Fで100%のスコアを記録し、実際のRustのバグも検出するLean 4モデルで、形式検証がオープンソースコードのセキュリティにおいて現実的な選択肢になりつつあることを示している。
mistral ai released leanstral 1.5, a free open-source model fine-tuned for the lean 4 language to formalize mathematical proofs and software verification. the model achieves 100% accuracy on the minif2f benchmark and solves 587 problems on putnambench, outperforming all open-source competitors and trailing only aleph prover. in practical tests, it scanned 57 open-source repositories and identified five previously unknown bugs, including a critical overflow in a rust library. this release signals that open-source models are now capable of handling complex formal verification tasks previously reserved for closed systems. available via hugging face and free api, leanstral 1.5 represents a shift toward verifiable, math-grade ai.