Home
Meituan Unveils 56B-Parameter LongCat, Setting New Benchmark in Open Source Mathematical Reasoning
As large language models expand into specialized domains, Meituan has delivered a standout solution that has captured the interest of both academia and industry.
On March 21, Meituan officially released LongCat-Flash-Prover, a large-scale mathematical reasoning model. This powerful system, featuring 567.7 billion parameters, utilizes an advanced MoE (Mixture of Experts) architecture and is specifically optimized for tackling highly complex formal mathematical proofs.

In top-tier benchmarks evaluating logical reasoning, LongCat-Flash-Prover has shown exceptional performance:
Record-Breaking Performance: It achieved a remarkable 97.1% score on the MiniF2F-Test, requiring only 72 reasoning attempts.
Conquering Hard Problems: It successfully solved 41.5% of the challenges in the PutnamBench dataset. Both results establish new global State-of-the-Art (SOTA) records.
To equip large models with the precision of a mathematician, Meituan implemented several critical technical innovations:
Reducing Hallucinations: A multi-stage verification process based on Abstract Syntax Trees (AST) was introduced, along with integration of the Lean4 formal language, effectively minimizing logical errors and "hallucinations" in AI deductions.
Advanced Training Methods: To address the instability often seen in long-horizon tasks within MoE models, Meituan deployed its proprietary HisPO algorithm. Combined with a theorem consistency check, this approach prevents reward hacking during reinforcement learning.
Optimized Architecture: With 560 billion total parameters, the model maintains a deep knowledge base, while the MoE structure ensures efficient and flexible inference.
Meituan has fully open-sourced the model and its source code on GitHub and Hugging Face.
With the release of LongCat-Flash-Prover
Related article
Meta Wants Your Next Gadget to Be Muse-Infused
Meta’s Muse, an AI assistant that handles travel bookings, form submissions, and shopping, has gained widespread adoption, yet a new initiative may especially intrigue developers and hardware enthusiasts.On Friday, Meta unveiled Muse Gadgets, an open
Wan-Streamer v0.2 delivers ultra-low latency AI chat
As modern technology advances rapidly, human-AI interaction is becoming more common. Traditional video calls often struggle with lag, delays, and audio-visual desynchronization. Wan-Streamer v0.2, the latest release from Tongyi Lab, overcomes these c
Meta Joins TSMC and Micron to Equip New Servers with Old Memory
As artificial intelligence adoption accelerates worldwide, next-generation DDR5 memory is experiencing a critical supply shortage. To avoid delays in AI data center deployments, Meta has partnered with industry leaders like TSMC and Micron to enginee
Related Special Topic Recommendations
Comments (0)
0/500
As large language models expand into specialized domains, Meituan has delivered a standout solution that has captured the interest of both academia and industry.
On March 21, Meituan officially released LongCat-Flash-Prover, a large-scale mathematical reasoning model. This powerful system, featuring 567.7 billion parameters, utilizes an advanced MoE (Mixture of Experts) architecture and is specifically optimized for tackling highly complex formal mathematical proofs.

In top-tier benchmarks evaluating logical reasoning, LongCat-Flash-Prover has shown exceptional performance:
Record-Breaking Performance: It achieved a remarkable 97.1% score on the MiniF2F-Test, requiring only 72 reasoning attempts.
Conquering Hard Problems: It successfully solved 41.5% of the challenges in the PutnamBench dataset. Both results establish new global State-of-the-Art (SOTA) records.
To equip large models with the precision of a mathematician, Meituan implemented several critical technical innovations:
Reducing Hallucinations: A multi-stage verification process based on Abstract Syntax Trees (AST) was introduced, along with integration of the Lean4 formal language, effectively minimizing logical errors and "hallucinations" in AI deductions.
Advanced Training Methods: To address the instability often seen in long-horizon tasks within MoE models, Meituan deployed its proprietary HisPO algorithm. Combined with a theorem consistency check, this approach prevents reward hacking during reinforcement learning.
Optimized Architecture: With 560 billion total parameters, the model maintains a deep knowledge base, while the
Meituan has fully open-sourced the model and its source code on GitHub and Hugging Face.
With the release of LongCat-Flash-Prover
Meta Wants Your Next Gadget to Be Muse-Infused
Meta’s Muse, an AI assistant that handles travel bookings, form submissions, and shopping, has gained widespread adoption, yet a new initiative may especially intrigue developers and hardware enthusiasts.On Friday, Meta unveiled Muse Gadgets, an open
Wan-Streamer v0.2 delivers ultra-low latency AI chat
As modern technology advances rapidly, human-AI interaction is becoming more common. Traditional video calls often struggle with lag, delays, and audio-visual desynchronization. Wan-Streamer v0.2, the latest release from Tongyi Lab, overcomes these c
Meta Joins TSMC and Micron to Equip New Servers with Old Memory
As artificial intelligence adoption accelerates worldwide, next-generation DDR5 memory is experiencing a critical supply shortage. To avoid delays in AI data center deployments, Meta has partnered with industry leaders like TSMC and Micron to enginee











