Meituan's LongCat-2.5 is a 1.6T-param model built for AI agents
A mixture-of-experts model with 48 billion active parameters, a 1-million-token context window and native image handling, aimed at long-running coding and design agents rather than chat.
Meituan released LongCat-2.5 on 2 November, a mixture-of-experts model with 1.6 trillion total parameters and roughly 48 billion active per token. It is built for long-horizon agentic work: sustained execution across terminals, browsers, GUIs, spreadsheets and design tools, with training that turned failures from real agent runs into reproducible tasks.
The model takes a 1-million-token context window and can output up to 128,000 tokens. Vision is handled natively rather than bolted on, with images processed as discrete tokens inside the same framework as text, which Meituan says is what lets it handle visual coding and design work directly rather than through a separate vision module. It is live now through Meituan’s own chat service and API, with weights due on GitHub and Hugging Face later.
Why it matters: a 1.6T-parameter model built around sustained agent sessions rather than single replies is a serious new option for anyone evaluating long-running coding or design agents, and it widens the field of large Chinese models competing directly on agentic work.
The caveat: Meituan has not published any third-party benchmarks alongside the launch, and the open-weight release has not landed yet, so none of its capability claims can be independently checked for now.