Mistral AI released Leanstral 1.5 on July 2, an open-source formal verification model that uses mathematical proof to verify that software behaves as intended. The 119-billion parameter model scored 100 percent on the miniF2F benchmark and discovered five previously unknown bugs while scanning 57 open-source repositories.

Formal verification sounds academic, but it solves a real problem: how do you know code is actually correct without running it endlessly in production? Leanstral uses the Lean 4 proof language to express code correctness as mathematical statements that either verify or fail.

How It Works

You feed Leanstral code properties you want to verify. The model formulates a Lean proof that either establishes correctness or identifies why the code doesn’t meet its specification. It’s not debugging—it’s mathematical proof that a system behaves as you designed it.

Mistral tested the approach by pointing Leanstral at open-source repositories and asking it to verify code properties. The model found five bugs no human reviewer had caught. That’s not a lab result. That’s production value.

Accessibility and Open Source

Mistral released Leanstral under the Apache 2.0 license. The model runs through a free API endpoint on Mistral’s platform, and full weights are downloadable from Hugging Face. Engineers can integrate it into their workflows immediately without vendor lock-in.

The model uses only 6 billion active parameters out of 119 billion total, balancing computational efficiency with reasoning capability. That design choice makes it practical for real codebases, not just toy examples.

Software verification is becoming important as AI helps write more code. If humans won’t read it, machines should at least verify it works as claimed.

When AI finds bugs in open-source code at scale, that’s a signal formal verification has left the lab.