Zhipu says GLM-5 runs on seven Chinese AI accelerator platforms

Zhipu says its GLM-5 model has been adapted to run on seven domestic Chinese AI chip platforms: Huawei Ascend, Moore Threads, Hygon, Cambricon, Kunlunxin, MetaX and Enflame.

SemiAnalysis reports the claim on X, citing a Zhipu paper that describes optimization from kernels up to inference frameworks and says the model has been served on these accelerators for many months. The post notes that the paper is almost a year old. SemiAnalysis's account is a secondhand report of Zhipu's claim, and the paper's performance results are not described in the reports.