Meituan releases trillion-parameter large model LongCat-2.0, the first trillion-parameter model to complete full-process training on a domestic computing cluster.

Deep Tide TechFlow News, June 30 - According to Meituan's official announcement, Meituan has officially launched its next-generation large model LongCat-2.0 and open-sourced it simultaneously. The model has a total of 1.6T parameters, making it the industry's first trillion-parameter model to complete full-process training and inference on a 50,000-card domestic computing cluster. It natively supports a 1M ultra-long context and focuses on code understanding, generation, and execution in Agentic Coding scenarios.

On the technical front, LongCat-2.0 adopts the LongCat Sparse Attention (LSA) mechanism, reducing long-text computation from quadratic to linear complexity. Through a zero-computation expert mechanism, it achieves token-level dynamic activation (33B~56B). It also introduces a MOPD architecture that integrates three expert capabilities: Agent, Reasoning, and Interaction. In terms of training efficiency, the team spent three years overcoming challenges in adapting domestic computing power, reducing the average daily failure rate by over 70%, increasing training MFU by 1.5 times, and achieving a stable daily throughput of over 1T tokens/day.

In performance benchmarks, LongCat-2.0 scored 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). On BrowseComp, it scored 79.9, reaching the level of cutting-edge closed-source models.