Leanstral 1.5: Proof Abundance for All - mistral.ai

2026-07-02 · Source: mistral.ai via Google News · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Emerging Technologies & Innovation · Depth: Expert, medium

Summary

Leanstral 1.5, a free Apache-2.0 licensed model from Mistral AI with 6B active parameters, significantly enhances formal verification capabilities. Released on July 2, 2026, it saturates miniF2F, solves 587 out of 672 PutnamBench problems, and achieves new performance records on FATE-H (87%) and FATE-X (34%). The model is trained using a three-stage process involving mid-training, supervised fine-tuning, and reinforcement learning with CISPO, utilizing multiturn and code agent environments. Beyond benchmarks, Leanstral 1.5 demonstrates practical utility by verifying complex code properties and uncovering 5 previously unknown bugs across 57 open-source repositories, including an overflow bug in the "datrs/varinteger" library. It also proved O(log n) time complexity for AVL trees. Fully open-sourced, it is available via Hugging Face and a free API endpoint.

Key takeaway

For Machine Learning Engineers or Software Engineers integrating formal methods, Leanstral 1.5 offers a powerful, cost-effective solution for proof engineering and code verification. You should consider deploying its free API or Hugging Face model to automate complex mathematical proofs or proactively identify subtle bugs in your Rust codebases, as demonstrated by its ability to find 5 new bugs. This can significantly enhance code reliability and reduce manual verification effort.

Key insights

Leanstral 1.5 significantly advances formal verification, making rigorous methods practical for real-world code and complex mathematics.

Principles

• Agentic RL improves long-horizon proof tasks.
• Test-time scaling boosts complex problem solving.
• Formal verification finds subtle, real-world bugs.

Method

Leanstral 1.5 undergoes mid-training, supervised fine-tuning, and reinforcement learning with CISPO, utilizing multiturn and code agent environments for proof and code verification.

In practice

• Use for verifying O(log n) complexity in AVL trees.
• Automate bug discovery in Rust codebases.
• Tackle IMO-level math challenges.

Topics

• Formal Verification
• Lean 4
• Reinforcement Learning
• Code Verification
• Bug Discovery
• Large Language Models

Code references

• mistralai/LeanstralSafeVerify
• mistralai/FLTEval
• datrs/varinteger
• oOo0oOo/lean-lsp-mcp

Best for: AI Architect, Research Scientist, CTO, AI Scientist, Machine Learning Engineer, Software Engineer

Related on AIssential

• How a 20-Person Startup Won Gold at the Math Olympiad—Tying With OpenAI & DeepMind (Tudor Achim, CEO of Harmonic) — Formal verification, powered by Lean, is crucial for building trustworthy and scalable AI systems capable of advanced mathematical reasoning.
• Verifying the Rust Standard Library — Large-scale, crowdsourced formal verification of the Rust standard library is feasible using multi-tool continuous integration.
• Process-Verified Reinforcement Learning for Theorem Proving via Lean — Lean proof assistant provides fine-grained, process-level verified feedback for reinforcement learning in theorem proving.
• With $1 Cyberattacks on the Rise, Durable Defenses Pay Off — Generative AI accelerates both cyberattacks and defenses, necessitating a shift to foundational security rather than reactive patching.
• Advancing Mathematics Research with AI-Driven Formal Proof Search — LLM-driven formal proof search, verified by Lean, can solve open mathematical problems and advance research.
• Introduction to Lean for Programmers — Mathematical propositions can be treated as types, and their proofs as programs, via the Curry-Howard correspondence.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by mistral.ai via Google News.