Meituan LongCat Unveils Open-Source Theorem Prover Model LongCat-Flash-Prover
On March 24, 2026, the Meituan LongCat team officially open-sourced a specialized deep learning model for mathematical formalization and theorem proving: LongCat-Flash-Prover. This model overcomes the limitations of large language models in rigorous logical reasoning by breaking down formal reasoning into three core capabilities: auto-formalization, proof sketching, and final proving. It represents a paradigm shift from "probabilistic answer prediction" to "verifiable logical proof."

Using a combined Tool-Integrated Reasoning (TIR) strategy, the model achieved a 97.1% pass rate on the MiniF2F-Test benchmark with only 72 reasoning steps, setting a new state-of-the-art record for open-source theorem provers. Its performance also significantly exceeded existing open-source models on challenging competition-level benchmarks like MathOlympiad-Bench and PutnamBench.

Technically, LongCat-Flash-Prover employs a TIR-based "hybrid expert iteration" framework. By integrating Lean4Server verification, semantic and theorem consistency checks, and legality verification against nine types of cheating behaviors, the model effectively addresses logical loopholes and code deception. During training, the team introduced a hierarchical masking strategy and token-level staleness control, greatly improving reinforcement learning stability under the Mixture-of-Experts (MoE) architecture.
As AI reasoning evolves from handling natural language ambiguity to working with verifiable formal languages, such theorem provers are moving beyond simple algorithmic benchmarks. They are becoming foundational infrastructure for core scientific research. This breakthrough signals an accelerating era where AI deeply participates in frontier mathematical exploration and automated document verification.
GitHub:
https://github.com/meituan-longcat/LongCat-Flash-Prover
Hugging Face:https://huggingface.co/meituan-longcat/LongCat-Flash-Prover
Report:
https://github.com/meituan-longcat/LongCat-Flash-Prover/blob/main/LongCat_Flash_Prover_Technical_Report.pdf
Related article
Stripe to Acquire OpenRouter as AI Model Routing Gains Momentum
Stripe has agreed to acquire OpenRouter, an AI model-routing platform that gives developers access to hundreds of models through a single interface. The deal adds model selection and routing to Stripe’s existing work around AI usage and token-based b
Former OpenAI Researcher Barrett Zoph Returns to Google to Support Gemini Development
Barret Zoph, who previously abruptly left and joined OpenAI due to a disagreement with Thinking Machines Lab CEO Miriam Mularati, has once again changed employers. Alphabet's Google announced that Zoph will return as Vice President of Research, bring
How Recursive Intelligence Raised $335M at $4B Valuation in 4 Months
The co-founders of startup Ricursive Intelligence seemed destined to be co-founders.Anna Goldie, CEO, and Azalia Mirhoseini, CTO, are so well-known in the AI community that they were among those AI engineers who “got those weird emails from Zuckerber
Related Special Topic Recommendations
Comments (2)
0/500
On March 24, 2026, the Meituan LongCat team officially open-sourced a specialized deep learning model for mathematical formalization and theorem proving: LongCat-Flash-Prover. This model overcomes the limitations of large language models in rigorous logical reasoning by breaking down formal reasoning into three core capabilities: auto-formalization, proof sketching, and final proving. It represents a paradigm shift from "probabilistic answer prediction" to "verifiable logical proof."

Using a combined Tool-Integrated Reasoning (TIR) strategy, the model achieved a 97.1% pass rate on the MiniF2F-Test benchmark with only 72 reasoning steps, setting a new state-of-the-art record for open-source theorem provers. Its performance also significantly exceeded existing open-source models on challenging competition-level benchmarks like MathOlympiad-Bench and PutnamBench.

Technically, LongCat-Flash-Prover employs a TIR-based "hybrid expert iteration" framework. By integrating Lean4Server verification, semantic and theorem consistency checks, and legality verification against nine types of cheating behaviors, the model effectively addresses logical loopholes and code deception. During training, the team introduced a hierarchical masking strategy and token-level staleness control, greatly improving reinforcement learning stability under the Mixture-of-Experts (MoE) architecture.
As AI reasoning evolves from handling natural language ambiguity to working with verifiable formal languages, such theorem provers are moving beyond simple algorithmic benchmarks. They are becoming foundational infrastructure for core scientific research. This breakthrough signals an accelerating era where AI deeply participates in frontier mathematical exploration and automated document verification.
GitHub:
https://github.com/meituan-longcat/LongCat-Flash-Prover
Hugging Face:https://huggingface.co/meituan-longcat/LongCat-Flash-Prover
Report:
https://github.com/meituan-longcat/LongCat-Flash-Prover/blob/main/LongCat_Flash_Prover_Technical_Report.pdf
Stripe to Acquire OpenRouter as AI Model Routing Gains Momentum
Stripe has agreed to acquire OpenRouter, an AI model-routing platform that gives developers access to hundreds of models through a single interface. The deal adds model selection and routing to Stripe’s existing work around AI usage and token-based b
Former OpenAI Researcher Barrett Zoph Returns to Google to Support Gemini Development
Barret Zoph, who previously abruptly left and joined OpenAI due to a disagreement with Thinking Machines Lab CEO Miriam Mularati, has once again changed employers. Alphabet's Google announced that Zoph will return as Vice President of Research, bring
How Recursive Intelligence Raised $335M at $4B Valuation in 4 Months
The co-founders of startup Ricursive Intelligence seemed destined to be co-founders.Anna Goldie, CEO, and Azalia Mirhoseini, CTO, are so well-known in the AI community that they were among those AI engineers who “got those weird emails from Zuckerber





Home






