Meituan releases LongCat-2.5: a multimodal agent model with 1.6 trillion parameters, trained on Chinese chips

The Chinese company Meituan, known for its food delivery app and other local services, released a new AI model on 2 November: LongCat-2.5. The model processes text and images at the same time and, according to Meituan, is built mainly for AI agents: programs that carry out a long task on their own, such as building a whole software project over many steps.

Anyone can try the model on the website longcat.ai, and developers can use it in their own software through an API. Meituan also wants to make the model freely available later through GitHub and Hugging Face, two sites where developers share code and AI models.

Large, but economical

LongCat-2.5 has 1.6 trillion parameters in total, the numbers in which an AI model stores its knowledge. For each piece of text it uses only about 48 billion of them: the model is made up of many specialised parts, and at each step only the parts that are needed take part. This makes the computing work lighter. Through the API, the model can process texts of about a million tokens at once. A token is a piece of a word; a million tokens is roughly a large pile of books.

The model turns images into the same kind of pieces as text, so it handles both in the same way. According to Meituan, this lets it, for example, build a web page from a design, or make pages with 3D images. To teach the model to program, Meituan had it practise on a large scale with complete software projects, rewarding it for what went well. The company turned mistakes that AI agents made in practice into new practice tasks.

Entirely on Chinese chips

According to Meituan, LongCat-2.5 was trained entirely on Chinese computing power. This matters because American export rules make it hard for China to get the best chips, for example from Nvidia. Meituan has done this before: in June it trained LongCat-2.0, which the company says was the first model with more than a trillion parameters to be both trained and used entirely on Chinese chips.

Chinese AI chips have less memory per card. That is why Meituan worked with Chinese chip makers to spread the work over the cards in a smarter way. As a result, during part of the training each card did 86 per cent more work than with the usual method, Meituan reports. When the model is in use, it also calculates its numbers in a more compact form (4 bit instead of 8 bit). In Meituan's tests this made the model about 40 per cent faster, without the answers getting worse.

Performance according to Meituan itself

Meituan says that LongCat-2.5 scores between Claude Opus 4.8 and Claude Opus 5 from the American company Anthropic on important well-known tests, and on some tests even higher than Opus 5. These are the company's own figures. Independent measurements still have to show how the model performs in practice.

Meituan does not want to earn money with the model by selling computing time. Its chief executive Wang Xing said at the end of August that the company does not want to become a "token factory": the AI is meant to help its own services. The model is the basis of Miaoshou, an AI tool for business owners that chains such as Guming, Xushengji and Yihetang already use to manage their shops, according to Meituan.

What does this mean for the road to the singularity?

AI models that carry out long tasks on their own are an important step towards AI that really takes over work. The fact that Meituan, by its own account, was able to train such a large model entirely on Chinese chips shows that there is more than one route to the computing power this needs. If independent tests confirm the figures, more companies and countries will join the top of AI, and that can speed up development further.

What is the technological singularity? Read the explanation →