Zhipu open-sources GLM-5.2, ranking first in Code Arena blind test

Zhipu AI announced today that it has open-sourced the GLM-5.2 model. In Code Arena, a blind testing system with millions of global users, the model ranked first among available models, surpassing closed-source models including Claude Opus 4.7.

GLM-5.2 is designed for long-context tasks, supporting a stable 1M context. It performs between Claude Opus 4.7 and 4.8 on multiple long-context benchmarks, making it the top open-source model. Its coding ability is on par with Claude Opus 4.8.

The model is already adapted to domestic computing platforms like Huawei Ascend, T-Head, Moore Threads, and Cambricon, reducing FLOPs per token by 2.9x. The upcoming Ascend 950 supernode is expected to become its powerful computing base in the second half of the year.