DgcityMeituan, the Chinese company best known outside Asia for its food-delivery app, has published the full weights of a 1.6-trillion-parameter artificial intelligence model under one of the software industry’s most permissive licenses, the latest sign that China’s consumer-internet giants are joining the open-source race led by dedicated AI labs.
The model, LongCat-2.5, was posted on the Hugging Face repository under an MIT license, meaning developers and companies anywhere can download, modify and use it for commercial purposes without paying royalties. The weights appeared on November 9, roughly a week after Meituan first introduced the system. The company retains its trademarks and patents, but otherwise places few conditions on use.
LongCat-2.5 is a native multimodal system built with a mixture-of-experts architecture, the same broad design used by most of today’s largest models. Of its 1.6 trillion total parameters, only about 48 billion are activated for any given token, which keeps running costs far below what the total size might suggest. It can process text, images and video together, converting visual information into discrete tokens handled alongside language, and it offers a switch between extended reasoning and direct-answer modes. Its context window extends to one million tokens, a feature aimed at long software-engineering sessions and complex multi-step agent tasks.
The model’s benchmark scores place it within reach of Western frontier systems rather than clearly ahead of them. On Terminal Bench, a test of agentic coding, LongCat-2.5 scored 88.4, essentially level with the leading GPT system’s 88.8 and just behind Claude Opus at 89.1, while DeepSeek’s latest model led the field at 90.6. That competitive positioning matters: Meituan is presenting the system as a practical workhorse, not a record-chaser.
Behind the release is a notable engineering claim. Meituan says both the training run and the large-scale deployment were built entirely on domestic AI accelerator clusters rather than the Nvidia chips that dominate the field, with pre-training spanning millions of accelerator-days across more than 35 trillion tokens without serious setbacks. To make the model easier to run on Chinese hardware, the team also produced a four-bit quantized version that it says lifts prefill throughput by about 43 percent over an eight-bit baseline. For a company whose core business is dispatching riders and matching restaurants with customers, the message is unmistakable: even delivery platforms now see frontier AI as core infrastructure.
