Meituan releases LongCat-2.5, a 1.6-trillion-parameter agent model trained entirely on Chinese chips
Meituan releases a 1.6-trillion-parameter agent model that never touched an Nvidia GPU
Meituan released the weights of LongCat-2.5, a 1.6-trillion-parameter mixture-of-experts model built for autonomous multi-step agent and coding tasks, trained entirely on Chinese hardware with no Nvidia GPUs. The model activates about 48 billion parameters per token, handles up to one million tokens via API, understands text, images and video, and is available under the MIT license with a self-hosting guide for a 194-chip Huawei Ascend Atlas A2 cluster. Meituan reports 88.4 on Terminal Bench v2.1 versus Claude Opus 5's 89.1 and GPT-5.6 Sol's 88.8, and 38.5 on the harder Terminal Bench v4.0 against Opus 5's 53.9.
Meituan's open-weight 1.6T-parameter agent model shows how far Chinese hardware-only training can go, with benchmark scores placed directly against Claude Opus 5 and GPT-5.6 Sol.
Source:The Decoder · the-decoder.com