Meituan LongCat-2.5: 1.6T Parameters, Only 48B Work Per Token
In short
- Meituan, the Chinese delivery giant, launched LongCat-2.5 on November 2: a multimodal model for long coding and visual tasks.
- It has 1.6 trillion parameters in total, but only about 48 billion work on each token. That is about 3%.
- You can use it now in Meituan’s chat and API. Price, license and benchmarks are not published yet, and the weights are promised for GitHub and Hugging Face.
A food delivery company just shipped a 1.6 trillion parameter model. That sounds like a joke, but the design is smart, and it matters for your bill.
What happened
On November 2, 2026, Meituan launched LongCat-2.5, according to TechNode. It handles text and images in one unified token framework, and it is aimed at extended coding tasks and visual work.
The model has 1.6 trillion parameters in total, with about 48 billion activated per token. It is available through Meituan’s online chat and its API platform. Code and weights are planned for GitHub and Hugging Face, but they are not out yet.
What we do not know: the price, the license, the context length, and any benchmark numbers. TechNode reports none of them, and I could not reach a primary Meituan page to check. Treat everything beyond the size and the access as unconfirmed.
How to use it
Do not move any production work yet. If you build with AI, put a hold on one test: take a real coding task you pay for today, run it through LongCat-2.5 in the chat, and compare the result with your current tool.
Do the math when the price lands. Say you spend €100 a month on coding APIs. Every 10% cheaper per token is €10 a month, or €120 a year. A model that works on only 3% of its weights per token has a real chance to be priced low, but that is a hope, not a fact. Wait for the number.
If the weights appear on Hugging Face, check the license first. Then you can host it yourself or use a provider that does.
What it means
Big sparse models are now the normal way to get frontier size without frontier cost. Only a small slice runs per token, so compute per answer stays modest, while memory for all weights is still huge. That is why running it yourself will not be cheap.
The other signal: AI is no longer only OpenAI, Google and Anthropic. Companies outside the usual list ship serious models, and that keeps prices under pressure. You win from that.
Ivan’s Take
Interesting, but empty for now. No price and no benchmarks means no decision. I would test it in the chat on one real task today, and decide on money only when the API price is public.
Source: TechNode
If it doesn’t save you time or money, you won’t see it here
- Facts, not hypeEvery price and number is checked against the official source
- Always the real priceYou see what it costs before you click
- I say when it’s weakEven when it’s OpenAI or Google
