TechPrufer
Sign inGet started

Blogs›Mistral AI›Formal Verification Models

Formal Verification Models

Formal Verification Models

1
posts
2026

Mistral AI introduces Leanstral 1.5, a 6B active parameter, Apache-2.0 licensed model for formal verification. It achieves state-of-the-art results on benchmarks like miniF2F, PutnamBench, FATE-H, and FATE-X. The model is trained using a three-stage process including mid-training, supervised fine-tuning, and reinforcement learning with CISPO, enabling it to perform agentic proof engineering and real-world code verification. Leanstral 1.5 has demonstrated the ability to uncover previously unknown bugs in open-source repositories.

2026

Leanstral 1.5: Proof Abundance for All

7/2/2026

This post introduces Leanstral 1.5, a new formal verification model. It details the three-stage training process (mid-training, supervised fine-tuning, reinforcement learning with CISPO) and the two RL environments used: a multiturn environment for theorem proving and a code agent environment for real-world code verification. The post also presents evaluation results on miniF2F, PutnamBench, FATE-H, FATE-X, and FLTEval, highlighting performance improvements and cost efficiency. Case studies on AVL tree time complexity verification and bug discovery in open-source repositories are included. Instructions for getting started with Leanstral 1.5 via Hugging Face, a free API, and Mistral Vibe are provided.