Meituan releases a 1.6-trillion-parameter agent model that never touched an Nvidia GPU
Meituan has released LongCat-2.5, a 1.6-trillion-parameter model designed for AI agents that was trained exclusively on Chinese hardware. The model achieved a score of 88.4 on Terminal Bench, placing it near the performance levels of Claude Opus 5. By demonstrating that such large-scale models can be trained and deployed using 194 Huawei Ascend chips, this development provides a viable alternative path for AI development in regions facing restrictions on access to high-end Nvidia hardware.
Covered by 1 source
- TThe Decoder↗Jonathan KemperNov 15