Meituan releases a 1.6-trillion-parameter agent model that never touched an Nvidia GPU

Why it matters

Meituan, the Chinese food delivery company, has publicly released the weights of LongCat-2.5, a mixture-of-experts model designed for autonomous agents working through long, multi-step coding tasks. The release is notable because the model was trained entirely on Chinese hardware, with no Nvidia GPUs involved, and its self-hosting guide describes a cluster built around Huawei Ascend accelerators. Meituan quantizes the model to a low-bit precision that those Chinese chips process natively, keeping accuracy close to the uncompressed version while boosting throughput. On agent benchmarks, LongCat-2.5 lands just behind top Western models like Claude Opus 5 on terminal tasks, though Meituan ran some evaluations in its own environment. The weights ship under a permissive MIT license, allowing commercial use, modification, and redistribution. The move signals that open-weight Chinese models trained on domestic hardware are closing the gap with leading US systems on common benchmarks.