Meituan Releases Trillion-Parameter Large Model LongCat-2.0, the First Trillion-Parameter Model to Complete Full-Process Training on a Domestic Computing Cluster

According to Meituan's official release, Meituan has officially launched the new generation large model LongCat-2.0 and open-sourced it simultaneously. The model features a total of 1.6T parameters, making it the industry's first trillion-parameter model to complete full-process training and inference on a domestic computing cluster of 50,000 cards. It natively supports 1M ultra-long context and focuses primarily on code understanding, generation, and execution in Agentic Coding scenarios. Technically, LongCat-2.0 adopts the LongCat Sparse Attention (LSA) sparse attention mechanism, reducing long text computation complexity from quadratic to linear; achieves token-level dynamic activation (33B~56B) via a zero-computation expert mechanism; and introduces the MOPD architecture to fuse three sets of expert capabilities: Agent, Reasoning, and Interaction. In terms of training efficiency, the team spent three years overcoming challenges in adapting to domestic computing power, reducing the monthly average daily failure rate by over 70%, increasing training MFU by 1.5 times, and achieving steady-state daily throughput exceeding 1T tokens/day. In terms of performance evaluation, LongCat-2.0 achieved a score of 59.5 on SWE-bench Pro, surpassing Gemini 3.1 Pro (54.2), GPT-5.5 (58.6), and Claude Opus 4.6 (57.3); and achieved a score of 79.9 on BrowseComp.