Meituan’s LongCat team has unveiled a massive 560-billion-parameter open-source model designed to bridge the gap between static knowledge and active problem-solving. Known as LongCat-Flash-Thinking-2601, this Mixture-of-Experts (MoE) system excels at agentic tasks, where an AI must autonomously use software tools and search the web to fulfill complex requests. What sets this model apart is its resilience in the face of noisy real-world data, such as broken links or contradictory information, which often cause standard AI systems to fail. To achieve this, the researchers systematically incorporated imperfections into the training process, forcing the model to develop robust recovery strategies rather than just memorizing perfect patterns. The technical backbone relies on an asynchronous reinforcement learning framework called DORA, allowing the model to learn efficiently across 10,000 different digital environments. Furthermore, a novel Heavy Thinking mode allows the model to spend more computational power during the actual response phase—a process called inference scaling. By expanding both the depth and width of its internal reasoning process, the model can tackle multi-turn interactions that require deep logical foresight and adaptive planning.

Meituan’s LongCat team has unveiled a massive 560-billion-parameter open-source model designed to bridge the gap between static knowledge and active problem-solving. Known as LongCat-Flash-Thinking-2601, this Mixture-of-Experts (MoE) system excels at agentic tasks, where an AI must autonomously use software tools and search the web to fulfill complex requests. What sets this model apart is its resilience in the face of noisy real-world data, such as broken links or contradictory information, which often cause standard AI systems to fail. To achieve this, the researchers systematically incorporated imperfections into the training process, forcing the model to develop robust recovery strategies rather than just memorizing perfect patterns. The technical backbone relies on an asynchronous reinforcement learning framework called DORA, allowing the model to learn efficiently across 10,000 different digital environments. Furthermore, a novel Heavy Thinking mode allows the model to spend more computational power during the actual response phase—a process called inference scaling. By expanding both the depth and width of its internal reasoning process, the model can tackle multi-turn interactions that require deep logical foresight and adaptive planning.