Key facts
- Meituan said LongCat-2.5 is a native multimodal model built with 1.6 trillion parameters.
- Meituan said the model was trained entirely on homegrown chips and is designed for long-range agent tasks.
- Meituan said it plans to use the model to improve efficiency across its core businesses.
Meituan unveiled LongCat-2.5, a native multimodal artificial intelligence model, on Monday, the company said. Meituan said the model was trained entirely on homegrown chips and was built with 1.6 trillion parameters, a measure of the model's size.
Meituan said the model is designed to handle long-range agent tasks, meaning work carried out over multiple steps by software acting with limited human input. The company said it plans to use the model to improve efficiency across its core businesses.
Caixin reported that the launch reflects Chinese technology companies' growing reliance on homegrown chips to advance their AI capabilities despite hardware limitations, and a focus on earning revenue from AI through existing business ecosystems rather than selling raw computing power. The report did not name the chips used or their supplier.