Leanstral 1.5: Proof Abundance for All

4 hours ago

Leanstral 1.5 is a free Apache-2.0 licensed model with 6B active parameters, offering a significant performance upgrade in formal verification.
It achieves state-of-the-art results on benchmarks: 100% on miniF2F, 587/672 on PutnamBench, 87% on FATE-H, and 34% on FATE-X.
The model undergoes a three-stage training process: mid-training, supervised fine-tuning, and reinforcement learning with CISPO, using multiturn and code agent environments.
It demonstrates strong test-time scaling on PutnamBench, solving more problems as token budget increases, from 44 at 50k tokens to 587 at 4M tokens.
Leanstral 1.5 excels in code verification, proving time complexity for AVL trees and uncovering 5 previously unknown bugs across 57 repositories tested.
The model is fully open-sourced and available via Hugging Face and a free API, with recommendations for setup using Mistral Vibe and Lean LSP MCP.

Hasty Briefsbeta