Mistral AI has released Leanstral 1.5, an open-source model built for formal verification in Lean 4. According to The Decoder, the model tops formal math benchmarks — but its most striking result reaches beyond pure mathematics.
Formal verification is the practice of proving, with mathematical rigor, that a piece of code or a mathematical statement does exactly what it claims. Lean 4 is a programming language and "proof assistant" designed for this: instead of testing software by running it on a few examples, you write a machine-checked proof that leaves no room for a hidden edge case. The catch is that writing these proofs is painstaking, which is why an AI model that can help generate them is significant.
The Decoder reports that Leanstral 1.5 does more than solve benchmark problems. When the model was pointed at 57 open-source code repositories, it surfaced five previously unknown bugs. In other words, it did not just prove tidy theorems in a controlled setting — it caught genuine defects in software that real people use.
That combination matters. Benchmark scores show a model is good at a test; finding real, previously undiscovered bugs shows it can do useful work in the messy world of production code. And because Mistral released Leanstral 1.5 as open source, researchers, developers, and companies can inspect it, run it themselves, and build on it without paying for access or trusting a closed system.
Formal verification has long been considered too slow and specialized for everyday software, reserved instead for high-stakes systems like aircraft controls or cryptography. Tools that lower the effort involved could gradually widen where that rigor is applied.
Why it matters: if AI models can make mathematically airtight code verification cheaper and more accessible, more of the software we rely on could be proven correct rather than merely tested — and Leanstral 1.5's real-world bug finds are an early sign that goal is becoming practical.