- Published on
A 1.6T-Parameter Open MoE Model with Native 1M Context and LongCat Sparse Attention
- Authors
- Name
- DeepHorus
Meituan has released LongCat-2.0, a massive Mixture-of-Experts language model with 1.6 trillion total parameters and 48 billion activated per token. The model supports a native 1-million-token context window, a significant leap for agentic coding tasks. Notably, both training and inference were conducted entirely on domestic AI ASIC superpods, reducing dependency on Nvidia hardware.
LongCat-2.0 succeeds LongCat-Flash, a 560-billion-parameter model released in 2025. The pretraining phase consumed over 35 trillion tokens and millions of accelerator-hours, with Meituan reporting zero rollbacks or loss spikes. This stability is particularly noteworthy given the less mature software ecosystem for non-Nvidia chips.
The architecture incorporates several innovations to manage computational costs. Zero-computation experts route simple tokens through a bypass, while a PID controller adjusts expert activation between 33 and 56 billion parameters. LongCat Sparse Attention uses three indexing methods to approximate linear scaling for long contexts. An N-gram embedding module with 135 billion parameters captures local token relationships, and the MOPD post-training pipeline fuses agent, reasoning, and interaction capabilities.
On benchmarks, Meituan claims LongCat-2.0 edges out GPT-5.5 on SWE-bench Pro with a score of 58.6 and matches Google's Gemini 3.1 Pro overall. However, on general agent benchmarks like FORTE and BrowseComp, it trails leading frontier systems. Independent verification from third-party evaluators has not yet been published.
The model's primary use cases include whole-repository reasoning, multi-step terminal tasks, and repository-level edits. These applications benefit from the 1M-token context window, which allows feeding entire codebases without summarization.
Industry observers note that LongCat-2.0 demonstrates China's growing capability in large-scale AI models using domestic hardware. If independent benchmarks confirm its performance, it could accelerate adoption of alternative chip ecosystems and challenge Nvidia's dominance in AI training.
Looking ahead, Meituan plans to release LongCat-2.0 as an open model, potentially fostering agentic coding applications. The success of its sparse attention and expert routing techniques may influence future model designs. However, questions remain about its general reasoning abilities compared to top-tier frontier models.
