Leanstral 1.5: An Open-Source Lean 4 Model Brings Formal Proof to Real Code

Why an Open-Source Theorem Prover Matters

For most of computing history, formal verification lived behind a wall. Proving that a program does what it claims, mathematically and without hand-waving, demanded specialists fluent in proof assistants like Lean 4, plus the patience to encode every lemma by hand. Leanstral 1.5, released by Mistral AI on July 2, 2026 under a permissive Apache-2.0 license, is a serious attempt to lower that wall. It is a frontier-grade theorem-proving model you can download, inspect, and run yourself, and I think its arrival marks a genuine inflection point for formal proof engineering.

The architecture is a study in efficiency. Leanstral 1.5 is a 119-billion-parameter mixture-of-experts model that activates only about 6 billion parameters per token, paired with a 256K-token context window. That combination lets it hold long proof states, intermediate goals, and library context in view at once, which is exactly what Lean 4 work demands. You are not feeding it a snippet and hoping; you are giving it the whole proof environment.

Saturating the Math Benchmarks

The benchmark results read like a milestone list. On miniF2F, the standard olympiad-style formal-math benchmark, Leanstral 1.5 effectively saturates the test, leaving little headroom for successors. On PutnamBench, a notoriously hard collection drawn from the Putnam competition, it formally solves 587 of 672 problems. It also posts state-of-the-art results on the FATE-H (87%) and FATE-X (34%) benchmarks, the latter being difficult enough that a third of the way through is itself a strong showing.

These are not results you fake with clever prompting. A Lean 4 proof either type-checks or it does not; the kernel is an unforgiving referee. When a model solves 587 Putnam problems in a formally verified way, those proofs are correct in the strongest sense we know how to express.

From Olympiad Math to Software Correctness

What elevates this release from impressive to important is that its authors pushed past competition math into everyday software. Using an automated pipeline, Leanstral 1.5 scanned 57 code repositories and flagged 11 genuine bugs, five of which had never been reported on GitHub. It also produced a formal proof that a real-world AVL-tree implementation maintains O(log n) complexity, turning a textbook complexity claim into a machine-checked guarantee about actual code.

That is the part worth dwelling on. Bug-finding tools are common; ones that reason at the level of formal proof and surface previously unknown defects are not. This is the difference between a linter that pattern-matches on suspicious code and a system that understands what the code is supposed to guarantee and checks whether it does.

Democratizing Rigorous Proof Engineering

Because Leanstral 1.5 is fully open and freely downloadable, its value compounds. Researchers can fine-tune it for domain-specific verification. Toolmakers can wire it into continuous-integration pipelines so proofs run alongside unit tests. Educators can put a capable Lean 4 assistant in front of students who would otherwise never touch a proof assistant. Open weights turn a demonstration into infrastructure.

I do not want to overclaim: formal verification will not replace testing overnight, and coaxing proofs out of any model still takes skill. But the trajectory is clear and encouraging. A rigorous mathematical guarantee that once required a specialist and weeks of effort is edging toward something a well-equipped engineering team can reach for on an ordinary Tuesday. When correctness becomes cheaper to prove, we build sturdier software. Leanstral 1.5 is a concrete, open step in that direction.

Sources: Mistral AI, July 2, 2026; MarkTechPost, July 3, 2026; The Decoder, July 2, 2026.