TechPrufer
Get started
Mistral AI
Mistral AI logoMistral AI · Feature trail

Formal Verification Models

1 post · Jul 2024 – Jul 2026

What this trail covers

Mistral AI introduces Leanstral, an open-source code agent for Lean 4, designed for formal verification and proof engineering. It leverages a sparse 6B active parameter architecture for efficiency and cost-effectiveness, outperforming larger open-source models and offering competitive performance against closed-source alternatives like Claude Opus at a significantly lower cost. Leanstral is integrated into Mistral Vibe and available via a Labs API, with weights released under Apache 2.0. Case studies demonstrate its ability to answer StackExchange posts about Lean changes and reason about programs by translating Rocq definitions to Lean and proving properties.

All Mistral AI trails
  1. Leanstral 1.5: Proof Abundance for AllLeanstral 1.5, a free Apache-2.0 licensed model with 6B active parameters (119B total), delivers significant performance upgrades in formal verification. It saturates miniF2F, solves 587/672 PutnamBench problems, and achieves state-of-the-art results on FATE-H (87%) and FATE-X (34%). The model is trained through a three-stage process: mid-training, supervised fine-tuning, and reinforcement learning with CISPO, utilizing both multiturn and code agent environments. It demonstrates practical proof engineering by uncovering 5 previously unknown bugs across 57 tested repositories. The post details evaluation benchmarks, code verification case studies (AVL Trees, bug discovery), and provides setup instructions for Mistral Vibe.
  2. Leanstral: Open-Source foundation for trustworthy vibe-coding | Mistral AIThis post introduces Leanstral, an open-source code agent specifically for Lean 4, focusing on formal verification and proof engineering. It highlights Leanstral's sparse 6B active parameter architecture, its efficiency and cost-effectiveness compared to larger models and closed-source alternatives, and its integration into Mistral Vibe and availability via a Labs API. The post includes detailed evaluations comparing Leanstral against other OSS models and the Claude family, showcasing its performance on realistic proof engineering scenarios (FLTEval). Case studies demonstrate its practical application in answering Lean-related StackExchange questions and translating/reasoning about programs from Rocq.
  3. MathΣtral | Mistral AIIntroduces Mathstral, a specialized model for STEM subjects, building on Mistral 7B. It achieves state-of-the-art reasoning in its size category on benchmarks like MATH (56.6%) and MMLU (63.47%). Mathstral can be further improved with additional inference computation, such as majority voting (68.37% on MATH) or using a strong reward model (74.59% on MATH with 64 candidates). It is an instructed model available for use or fine-tuning, with weights hosted on HuggingFace and integration with mistral-inference and mistral-finetune.