Meituan is the company most people know for food delivery in China. On November 2, that same company released a new AI model called LongCat-2.5, and the pitch is not about chat. It is about work that runs for a long time: coding jobs that stretch across terminals, browsers, spreadsheets, and design tools, with pictures in the mix too.
The numbers, in plain words
LongCat-2.5 has 1.6 trillion parameters in total, but only about 48 billion of them switch on for each token it processes. Think of it as a very large team where only a small crew handles any single word. That design is called mixture-of-experts, and it is how the model stays fast while staying big.
It also has a 1 million token context window, which means it can hold a whole codebase or a long multi-step job in memory at once, and it can write up to 128,000 tokens in one output. On Meituan's own platform, the price right now is 30 cents per million input tokens and 1.20 dollars per million output tokens, listed as a limited-time discounted rate.
What it is built to do
Meituan positions the model for long-horizon agent work. In practice that means jobs like these:
Writing and fixing code across a full project, one file at a time
Working through a browser or a desktop screen step by step
Turning a written brief into visuals, like an editable 3D scene or an SVG graphic
Carrying a task across spreadsheets and design tools without losing track
It also speaks both the OpenAI and the Anthropic request formats, so it can drop into tools like Claude Code by changing a base URL rather than rewriting a client.
Two honest caveats
First, the parameter count is disputed. One early report said the 1.6 trillion and 48 billion figures were carried over from the LongCat-2.0 model card, so treat the headline size with a little care until Meituan's own technical write-up settles it.
Second, the benchmark scores are Meituan's own, not independently checked yet. The company reports 88.4 on Terminal Bench v2.1, up from 70.8 for LongCat-2.0, but outside testers have not confirmed those numbers so far. The weights are not open yet either. Meituan says GitHub and Hugging Face releases are planned, but no date is set, so for now you use it through its chat service and API.
So here is what I want to know. If a model can sit with a project for hours instead of minutes, does that change what you would hand to it? Would you trust it with a full coding job, or would you still keep the hard parts for yourself? Tell me in the comments.
Cover image: AI-generated.
