LongCat-Flash-Prover

Overview

LongCat-Flash-Prover is a 560-billion-parameter Mixture-of-Experts (MoE) model released in March 2026, purpose-built for formal mathematics and theorem proving. Unlike general-purpose coding or math models that aim only for a correct numeric answer, Flash-Prover produces machine-verifiable proofs in Lean4 — a formal language where every reasoning step is mechanically checked by a compiler. The model decomposes the theorem-proving task into three atomic capabilities: Auto-Formalization (translating informal problems into formal Lean4 statements), Sketching (drafting a high-level proof strategy with lemma decomposition), and Proving (filling in the detailed proof steps with compiler-guided self-correction). With TIR, the model iterates between proof attempts and compiler feedback until a valid proof is found or the attempt budget is exhausted. Flash-Prover achieves a 97.1% pass rate on MiniF2F-Test with only 72 inference attempts per problem — a new open-source state-of-the-art. On competition-level benchmarks, it solves 46.7% of MathOlympiad-Bench (180 attempts) and 41.5% of PutnamBench (118 attempts). Auto-formalization accuracy reaches 100% on MiniF2F-Test and ProofNet, meaning the model makes zero syntax errors when translating natural-language math problems into Lean4.

Architecture

Flash-Prover extends the 560B ScMoE architecture through integration with Lean4 as a formal verification tool within the TIR loop. The Auto-Formalization model generates N candidate formal statements from an informal problem, each checked by the Lean4 server for syntax and a semantic consistency judge for meaning. The Sketching model produces lemma-style proof strategies that decompose difficult goals into sub-goals — the publisher reports that sketching improves proof success by approximately 10% under the same compute budget. The Proving model then completes each lemma using TIR: generate proof code → submit to Lean4 → use compiler errors to guide the next attempt. The multi-turn correction loop continues until either a valid proof is found or the attempt budget is exhausted. During RL training, the HisPO (Hierarchical Importance Sampling Policy Optimization) algorithm addresses stability challenges through: sequence-level masking (filtering trajectories with excessive train-inference mismatch), token-level masking (removing unstable tokens), and token-level clipping (controlling stale policy updates). Without these, the publisher reports that MoE models exhibit reward hacking — exploiting compiler pass/fail signals through early exits, fake axioms, and macro-based bypass tricks.

Training

Training uses a two-phase hybrid-experts iterative framework. In the cold-start phase, an early Auto-Formalizer (ATF-32B) synthesizes formal statements, and LongCat-Flash-Thinking-2601 generates proof trajectories with Lean4 feedback. These are filtered, de-duplicated, and sampled for diversity before SFT of a unified model. In the iteration phase, the current model becomes the next expert, synthesizing fresh trajectories per task and per capability, mixed with general-purpose data to preserve reasoning utility beyond formal math. Multiple rounds of SFT + RL with HisPO stabilization produce the final Prover model.

Key Benchmarks

Publisher-reported scores. Not independently verified by this site. Descriptions summarize what each benchmark measures.

Deployment

Model weights on HuggingFace (meituan-longcat/LongCat-Flash-Prover) and GitHub (meituan-longcat/LongCat-Flash-Prover) under MIT License. GitHub repo includes technical report PDF, inference code, and Lean4 integration documentation. Requires a GPU with sufficient VRAM for the 560B MoE model plus an active Lean4 installation for proof verification. Tree Search mode for unlimited-budget exploration adds ~3.1% average improvement across benchmarks.

Capabilities

Described in the cited primary documentation. Not an independent performance evaluation.

• Native formal reasoning in Lean4 — machine-verifiable proofs.
• Auto-Formalization: translates informal math statements into Lean4.
• Sketching: drafts lemma-style high-level proof plans.
• Proving: completes formal proof details with verifier feedback.
• Tool-Integrated Reasoning (TIR): multi-turn self-correction loops.
• HisPO algorithm stabilizes MoE RL training on long-horizon tasks.
• Theorem consistency and legality detection prevent reward hacking.
• Tree Search mode for unlimited-budget exploration.

Access & License

MIT License. Model weights and inference code open-sourced.

Sources

• LongCat-Flash-Prover (arXiv:2603.21065)

  Primary source · Published 2026-03-22 · Accessed 2026-08-02

  560B MoE, HisPO algorithm, hybrid-experts iteration. 97.1% MiniF2F (72 attempts), 41.5% PutnamBench, 70.8% ProverBench.

• LongCat-Flash-Prover on HuggingFace

  Primary source · Accessed 2026-08-02

  Model weights. MIT License.

• LongCat-Flash-Prover GitHub Repository

  Primary source · Accessed 2026-08-02

  Code, technical report PDF, Lean4 integration docs.

FAQ

Where can the primary source for LongCat-Flash-Prover be found?

The official GitHub repository (meituan-longcat) and HuggingFace page contain the primary documentation, model weights, and license for LongCat-Flash-Prover.

Is LongCat-Flash-Prover available through the LongCat API?

LongCat-Flash-Prover is documented in the meituan-longcat GitHub organization. Refer to the publisher API documentation at longcat.chat for current API availability.

Is this page an official LongCat or Meituan page?

No. This page is published by an independent third-party site. It is not affiliated with, endorsed by, or sponsored by LongCat or Meituan.

How often is this page updated?

This page was last verified on 2026-08-02. Content is reviewed when new publisher documentation or model releases become available.