Mistral Releases Leanstral 1.5, an Apache-2.0 Lean 4 Code Agent Solving 587 of 672 PutnamBench Problems

Mistral AI released Leanstral 1.5, an open model under a permissive Apache-2.0 license with 6 billion active parameters, purpose-built as a code agent for the Lean 4 theorem prover. It delivers state-of-the-art formal-verification results — 87% on FATE-H and 34% on FATE-X — and solves 587 out of 672 PutnamBench problems, a benchmark of competition-level mathematics.

The model is fully open-sourced and available via Hugging Face and a free API, aligning with Mistral's positioning as the champion of open, customer-controlled AI. Mistral's official account emphasized this week that "the value created from their data, workflows, feedback, and models accrues to them rather than to model providers" — a direct jab at the closed-model gating dominating the news cycle.

Formal verification is a strategically important niche: as AI-generated code proliferates and AI-driven vulnerability discovery spikes (see the June CVE surge), the ability to mathematically prove code correctness becomes more valuable. A small, open, Apache-licensed model that's competitive at Lean 4 lowers the barrier for researchers and engineers to adopt formal methods.

The release lands during an "open-source AI summer" push — Hugging Face's Clément Delangue teased exactly that, and Chinese open models like GLM-5.2 and Gemma 4 are surging. Watch whether Leanstral 1.5 gets integrated into verification toolchains and how it compares to DeepSeek and specialized proof models.