China's Meituan releases LongCat-2.5, a giant AI model trained on Chinese-made chips
The food-delivery company's model has 1.6 trillion parameters but uses only about 48 billion for each word, and it is built to work through long coding jobs on its own.
Meituan, the Chinese company best known for food delivery, released a new AI model Monday called LongCat-2.5, according to its technical blog. It is built for long, multistep jobs such as building a whole software project, and it can understand images as well as text.
What the numbers mean
An AI model's parameters are the internal settings it learns during training; more usually means more capable but costlier to run. LongCat-2.5 has 1.6 trillion, but uses only about 48 billion for each token, a word or piece of a word. That is because it is a "mixture of experts" design, which sends each piece of text to a few specialized sub-networks instead of the whole model. Many large Chinese models, including DeepSeek's, use the same approach to save computing power.
Meituan said it trained the model with reinforcement learning, in which the AI tries real coding tasks and is rewarded for working results. It also turned real failures by AI agents into new practice tasks.
Built on Chinese chips
The company said it optimized training for domestic Chinese AI chips, each with less than 80 gigabytes of memory, as U.S. export rules limit China's access to top Nvidia chips. Running the model with numbers stored in just 4 bits, a compression trick, raised its speed about 43% over an 8-bit version, Meituan said.
Meituan compared LongCat-2.5 with OpenAI's GPT-5.5 and Anthropic's Claude Opus 5 on its own tests, which have not been independently checked. Its blog did not say whether the model can be freely downloaded.
Source: Meituan technical blog (in Chinese), Nov. 2, 2026.