Meituan has released the weights for LongCat-2.5, an agent-focused model trained entirely on Chinese hardware. The release gives developers access to a very large mixture-of-experts system designed for long, multi-step work, especially software tasks.

LongCat-2.5 contains 1.6 trillion parameters but activates roughly 48 billion for each token, limiting the computation used for any one response. Through Meituan’s API it accepts up to one million tokens of context and can produce as many as 128,000 output tokens. It processes text, images and video, responds in text, and lets users turn reasoning on or off.

Meituan says it trained the model without Nvidia GPUs and shows a sample deployment using 194 Huawei Ascend chips. Its reinforcement-learning data included real tasks derived from GitHub pull requests and recreated agent failures; training environments had no internet access, Git history or cached files to reduce shortcuts. The company reports 88.4 on Terminal Bench, close to Claude Opus 5, but vendor-reported benchmark results do not by themselves show how reliably the model performs in production. The release is also a notable demonstration that a frontier-scale open-weight agent model can be built and served on a non-Nvidia stack.