One-sentence summary
Meituan's LongCat-2.5 (Nov 2) has 1.6T params, ~48B active per token, native multimodal 1M context, and was trained end-to-end on domestic Chinese compute, extending the domestic-chips-domestic-models push.
Background
On November 2, 2026, Meituan launched LongCat-2.5. Back in June it completed full training and inference of the trillion-parameter LongCat-2.0 on domestic compute; LongCat-2.5 builds on that.
Model Specs
LongCat-2.5 has 1.6T total parameters in a MoE architecture, activating about 48B per token, with native 1M-token context and image understanding for long-horizon agent tasks.
What Domestic Chips Mean
Meituan stresses training was done entirely on domestic compute, meaning the compute stack no longer relies on overseas GPUs, offering another case of trillion-parameter training on domestic chips.
Where It Lands
LongCat-2.5-Preview targets long-horizon tasks across browsers, terminals, desktop software, spreadsheets and design tools, strengthening vision and agentic coding.