Mistral
8/10 signal
Leanstral 1.5: Proof Abundance for All
reasoningagentictool-use
What happened
Mistral AI released Leanstral 1.5, an open-source (Apache-2.0) MoE model with 119B total and 6B active parameters optimized for formal verification in Lean 4. It achieves state-of-the-art results on math and verification benchmarks, solving 587/672 PutnamBench problems and hitting 87% on FATE-H. The model was trained using mid-training, SFT, and reinforcement learning with CISPO in a multi-turn environment where it iteratively refines proofs based on Lean compiler feedback. In real-world testing, it uncovered 5 previously unknown bugs across 57 repositories.
Why it matters
It demonstrates the viability of using compiler-feedback reinforcement learning to train open-source models that can formally verify code and find real-world bugs.
The take
Leanstral 1.5 is a prime example of how compiler-in-the-loop reinforcement learning can create highly specialized, hyper-reliable reasoning agents. By training the model to interact directly with the Lean compiler, Mistral has bypassed the reliability limitations of standard LLMs for code verification. This is a blueprint for how we will build future coding agents: tight execution-feedback loops during both training and inference.
Do this
If you are building coding agents or automated QA pipelines, evaluate Leanstral 1.5 on Hugging Face or Mistral's API to see how compiler-driven reasoning can be integrated into your workflows.
Don't read this site daily. Get it in your inbox.
The daily brief and Sunday deep dive — distilled, scored, and opinionated. For builders only.