Leanstral 1.5

Leanstral 1.5, published by Mistral AI on 2 July 2026, is a code agent for Lean 4 — the proof assistant in which mathematical arguments and program properties are written so that a machine, not a reviewer, checks whether they hold. That framing matters: unlike a chatbot asked to do mathematics, Leanstral's output is either accepted by the Lean kernel or it is not, which removes the usual question of whether the model is bluffing. The model is a mixture-of-experts with 119 billion total parameters and roughly 6 billion active per token, released under an Apache-2.0 licence as mistralai/Leanstral-1.5-119B-A6B on Hugging Face and served free through Mistral's Labs API. On benchmarks it saturates miniF2F at 100% on both validation and test sets, solves 587 of 672 PutnamBench problems, reaches 87% on FATE-H and 34% on the harder FATE-X, and lifts FLTEval from 28.9% at a single attempt to 43.2% across eight. Mistral puts the cost at roughly four dollars per PutnamBench problem solved. The most telling result is not a benchmark score but test-time scaling: raising the token budget from 50k to 4M lifts the number of Putnam problems solved from 44 to 587, meaning the model trades compute for correctness in a way few systems do so cleanly. Mistral also reports that Leanstral found five previously unknown bugs while scanning 57 open-source repositories. An earlier Leanstral endpoint (labs-leanstral-2603) was retired on 30 June 2026 and this release replaces it.

▸Videos

No videos yet.

▸Reviews