Mistral AI has just released Leanstral 1.5, a 119B parameter Mixture-of-Experts (MoE) model designed to bridge the gap between probabilistic guessing and mathematical certainty. By combining the creative power of large language models with Lean 4—a formal verification language—Mistral is pushing for a future of "proof abundance."
Reinforcement Learning with a Compiler Touch
What makes Leanstral 1.5 special isn't just its size; it’s how it was trained. Mistral used two distinct reinforcement learning environments where the model is forced to interact directly with the Lean compiler. In a multiturn setup, the model attempts a proof, receives feedback on its errors, and iterates until it finds a solution. This isn't just advanced autocomplete; it's a digital mathematician that learns from its own logical mistakes in real-time. With only 6.5B active parameters during inference, it’s surprisingly efficient for its weight class.
Crushing the PutnamBench
The benchmarks are, frankly, impressive. Leanstral 1.5 managed to solve 587 out of 672 problems on the PutnamBench, a benchmark notoriously difficult for both humans and machines. It utilizes a technique called "context compaction" to handle long, complex proofs without losing the logical thread of the argument. This allows the model to build auxiliary lemmas and persist through long-form reasoning tasks that would typically crash a standard LLM's context window.
The Future of Verified Logic
While "good enough" code has powered the web for decades, the stakes for critical infrastructure and smart contracts are rising. We are entering an era where logic needs to be mathematically provable. By releasing Leanstral 1.5 under an Apache 2.0 license, Mistral is making these high-level formal verification tools accessible to the open-source community. It suggests a world where bug-free, verified code isn't just a luxury for NASA—it’s a standard generated at scale for everyone.
Sources
Media
