One-sentence summary
Common questions on Meituan's Nov 2 LongCat-2.5: how 1.6T params activate, what 1M context means, what domestic chips mean, and how to use the Preview.
What is LongCat-2.5?
LongCat-2.5 is Meituan's next-generation LLM, launched November 2, 2026, a native multimodal model for long-horizon agent tasks with 1.6T total parameters.
How are 1.6T params activated?
LongCat-2.5 uses a MoE architecture, activating about 48B parameters per token; not all 1.6T run at once.
What does 1M-token context mean?
It means the model can handle up to about 1M tokens of context per pass, suiting long documents, long tasks and multi-step agent work.
What do domestic chips mean?
LongCat-2.5 was trained entirely on domestic compute, so the training and inference stack does not rely on overseas GPUs.
How do I use it now?
LongCat-2.5-Preview is live, targeting long-horizon tasks across browsers, terminals, desktop software, spreadsheets and design tools.
Sources
- Meituan: Meituan launches LongCat-2.5