NewsBite analysis · AI, Programming
The Decoder
Mistral's open-source Leanstral 1.5 aces formal math benchmarks and catches real bugs in code
Mistral AI has introduced Leanstral 1.5, an open-source model focused on formal verification in the Lean 4 programming language. The model excelled in various mathematical benchmarks, achieving 100% on miniF2F and strong results on additional tests like PutnamBench. Notably, it also identified five real bugs across 57 open-source repositories during practical evaluations. Leanstral aims to enhance both mathematical proof verification and software correctness.
Act nowConfidence: HighShort term horizon
Published 4 Jul 2026, 08:10 UTC1 min read
Visit original article ↗Lead image — The Decoder
Summary
TL;DR
- Leanstral 1.5 is an open-source model under Apache 2.0.
- Achieved 100% on miniF2F formal math benchmark.
- Solved 587 problems on PutnamBench from 672 presented.
- Scored 87% and 34% on the FATE-H and FATE-X algebra benchmarks.
- Identified five previously unknown bugs in open-source code.
Interesting facts
- The model was trained using mid-training, supervised fine-tuning, and reinforcement learning.
- Leanstral also supports bug detection beyond mathematical tasks.
Coverage gaps