TLDR: LongCat-Flash-Thinking is an efficient 560-billion-parameter open-source Mixture-of-Experts (MoE) AI model from Meituan. It achieves state-of-the-art performance in complex reasoning tasks like mathematics, coding, and agentic AI, thanks to a unique training process involving long Chain-of-Thought cold-start, a domain-parallel reinforcement learning scheme, and an industrial-scale asynchronous RL framework called DORA. The model significantly improves efficiency, reducing token consumption in agentic tasks while maintaining accuracy, and excels in formal theorem proving.
The Meituan LongCat Team has introduced a groundbreaking artificial intelligence model called LongCat-Flash-Thinking, an efficient 560-billion-parameter open-source Mixture-of-Experts (MoE) reasoning model. This new model is designed to push the boundaries of AI capabilities, particularly in complex reasoning tasks, and has achieved state-of-the-art performance among open-source models.
At its core, LongCat-Flash-Thinking boasts advanced capabilities cultivated through a meticulously designed training process. This journey begins with a ‘cold-start’ using extensive Chain-of-Thought (CoT) data, which significantly enhances the model’s foundational reasoning potential. This initial phase equips the model with specialized skills in both formal reasoning (like mathematical proofs) and agentic reasoning (problem-solving with tools).
A Novel Training Approach
A key innovation in the model’s development is its domain-parallel training scheme. This approach allows for the optimization of distinct domains, such as STEM (Science, Technology, Engineering, and Mathematics), Code, and Agentic tasks, to happen independently. After these specialized ‘expert’ models are trained, they are then fused into a single, highly capable model that performs exceptionally across all these areas.
Powering this entire process is the Dynamic ORchestration for Asynchronous rollout (DORA) system. DORA is a large-scale Reinforcement Learning (RL) framework that delivers a remarkable training speedup, more than threefold compared to traditional synchronous methods, even when utilizing tens of thousands of accelerators. This infrastructure is crucial for making such advanced and large-scale AI training feasible.
Core Contributions and Efficiency
The LongCat-Flash-Thinking project highlights three main contributions. Firstly, the Domain-Parallel RL Training and Fusion Methodology addresses the instability often found in traditional mixed-domain RL training. By decoupling optimization across STEM, coding, and agentic tasks, and then fusing the resulting domain-expert models, it creates a final model that excels across all specialties.
Secondly, the Pioneering Industrial-Scale RL Infrastructure, the DORA system, provides a robust backbone for this extensive RL training. Its asynchronous architecture not only speeds up training but also enables stable operation on a massive scale. DORA includes novel features like elastic colocation and efficient KV-cache reuse, which are vital for its performance.
Thirdly, the model demonstrates Broad and Efficient Advanced Reasoning. For agentic capabilities, a dual-path reasoning approach selects high-value training queries that benefit most from tool integration, complemented by an automated pipeline for constructing high-quality, tool-augmented reasoning trajectories. This leads to significant efficiency gains; for instance, on the AIME-25 benchmark, LongCat-Flash-Thinking reduces average token consumption by 64.5% (from 19,653 to 6,965) without sacrificing accuracy. For formal reasoning, an expert-iteration pipeline integrated with a Lean4 server synthetically generates verified proofs, instilling a capability often absent in other large language models.
Also Read:
- MapCoder-Lite: Empowering Small Language Models for Advanced Code Generation
- Roster of Experts: Enhancing LLM Predictions Through Internal Diversification
Performance Across Benchmarks
LongCat-Flash-Thinking has been rigorously evaluated across a diverse set of benchmarks, including general QA, alignment, mathematical reasoning, general reasoning, coding, agentic tool use, formal theorem proving, and safety. The results show that the model consistently achieves superior performance, often outperforming other open-source models and competing strongly with leading closed-source models, despite having fewer activated parameters.
For example, it achieves a highly competitive score of 99.2% on the MATH500 benchmark and demonstrates outstanding proficiency in olympiad-level mathematics. In coding, it scores 79.4% on LiveCodeBench, surpassing all listed open-source models. Its formal theorem proving capabilities are particularly strong, achieving a pass@1 score of 67.6% on MiniF2F-Test, significantly outperforming its closest open-source competitor. Furthermore, the model exhibits state-of-the-art performance in safety benchmarks, effectively refusing harmful or unethical queries across various categories.
The Meituan LongCat Team hopes that by open-sourcing LongCat-Flash-Thinking, it will further advance research in reasoning models, especially in high-quality data strategies, efficient RL training, and native agentic reasoning. For more in-depth technical details, you can read the full research paper here.
