Meituan Open-Sources LongCat-Flash-Prover Large Model: 560 Billion Parameters Set New SOTA Records in Math Proofs

Meituan recently open-sourced the LongCat-Flash-Prover large model, which uses a 567.7 billion-parameter Mixture of Experts (MoE) architecture and focuses on solving complex mathematical proof problems. It introduces a hybrid expert iteration framework to generate high-quality formal reasoning trajectories, and eliminates "hallucinations" by integrating Lean4 and multi-stage strict verification processes.

In training, the model uses the hybrid expert framework to generate cold-start data, adopts the HisPO algorithm in the reinforcement learning phase to stabilize long-term task training, and adds theorem consistency and legitimacy detection mechanisms to prevent reward hacking.

Benchmark tests show that the model achieved 97.1% on MiniF2F-Test (only 72 reasoning attempts) and solved 41.5% of problems on PutnamBench (118 attempts), both setting new SOTA records.