Meituan has officially announced the release and open-sourcing of its cutting-edge trillion-parameter large model, LongCat-2.0. This groundbreaking model has undergone comprehensive training and inference on a 50,000-card domestic computing cluster, setting a new industry benchmark as the first trillion-parameter model to accomplish this milestone. LongCat-2.0 is equipped with a staggering 1.6 trillion parameters, boasting an average activation of around 48 billion and a dynamic range spanning from 33 billion to 56 billion, providing native support for ultra-long 1M context.

Following the release of its preview version on the OpenRouter platform, LongCat-2.0 swiftly ascended to the top three in monthly call volume globally, establishing itself as one of the most sought-after Agent models among developers worldwide. This achievement underscores a significant leap forward in domestic computing capabilities for large-scale cluster training.

Since 2023, the LongCat team has dedicated three years to successfully surmounting fundamental challenges, including operator adaptation, communication optimization, and distributed stability. By developing proprietary deterministic operators and implementing elastic recovery mechanisms, the team has slashed the average daily failure rate by over 70%, achieving a steady-state daily throughput exceeding 1 trillion tokens.

In terms of architectural innovation, LongCat-2.0 has been meticulously crafted to excel in real-world Agentic Coding tasks. It incorporates a Sparse Attention Mechanism (LSA), which effectively reduces long-text computational complexity to a linear scale. Additionally, by leveraging the Zero-Computation Expert Mechanism and MOPD multi-expert fusion architecture, LongCat-2.0 enables Token-level dynamic activation, demonstrating exceptional performance in complex office scenarios such as code comprehension, mathematical reasoning, and long-range information retrieval.

In authoritative programming evaluations, such as SWE-bench Pro, LongCat-2.0 has outperformed competitors like GPT-5.5 and Claude Opus4.6, providing robust support for the closed-loop implementation and application restructuring of enterprise-level AI Agents.