Meituan Releases 1.6T Open MoE Model LongCat-2.0

Meituan released LongCat-2.0, a large-scale open Mixture-of-Experts (MoE) language model, on an unspecified date. This model boasts a total of 1.6 trillion parameters, with approximately 48 billion parameters activated per token. LongCat-2.0 is specifically designed for agentic coding tasks, including code understanding, generation, and execution within agent workflows. Two key features distinguish this release: a native 1-million-token context window and the fact that both training and serving were conducted entirely on domestic AI ASIC superpods, indicating a move towards self-sufficiency in advanced AI infrastructure.

LongCat-2.0 represents Meituan's next-generation trillion-parameter open model, succeeding LongCat-Flash, a 560 billion parameter model introduced in 2025. The architecture was developed with the primary objective of enabling reliable and efficient agentic coding. The pretraining process involved over 35 trillion tokens and spanned millions of accelerator-hours. Meituan reported no rollbacks or significant loss spikes during this extensive training run, a notable achievement, particularly on non-Nvidia hardware where tooling maturity can be a challenge.

The model's architecture incorporates four key innovations to manage the cost associated with its massive scale. These include zero-computation experts, where simple tokens like punctuation are processed by a minimal-compute expert, while complex tokens engage more expert capacity. A PID controller dynamically adjusts expert bias to maintain an average activation within a specified range, resulting in a dynamic activation window of 33 billion to 56 billion parameters, rather than a fixed cost. The MoE backbone utilizes a shortcut-connected design (ScMoE) to enhance throughput.

Furthermore, LongCat-2.0 features LongCat Sparse Attention (LSA), an evolution of DeepSeek Sparse Attention (DSA). LSA addresses the quadratic scaling of standard attention with context length by selectively attending to the most relevant tokens, thereby approximating linear scaling. This is achieved through three orthogonal indexing methods: streaming-aware indexing, which consolidates fragmented memory reads into contiguous blocks; cross-layer indexing, which reuses attention saliency across adjacent layers; and hierarchical indexing, which employs a coarse-to-fine, two-stage approach.