Ascend platform now supports GLM-5.2 on day zero, boosting China's AI computing ecosystem

Huawei's Ascend AI developer account announced that the Ascend platform now supports Zhipu AI's GLM-5.2 model from day zero. It's already running on the A3 series for both single and dual server setups, with optimized inference for the model's MOE architecture. They've used tricks like fused operators and communication-computation pipelining to boost efficiency.

Zhipu AI released GLM-5.2 as open source on June 17. It's built for long tasks with a 1M context window. On Code Arena, a coding benchmark, GLM-5.2 ranked first among all available models worldwide. Its coding skills are top among open-source models, matching Claude Opus 4.8.

GLM-5.2 is already adapted for several Chinese AI chips, including Ascend, T-Head, and Cambricon, from day zero. It cuts FLOPs per token by 2.9 times at 1M context. The upcoming Ascend 950 supernode later this year is expected to boost its computing power even more.