GLM-5.2 open-sourced: tops Code Arena, built for long-horizon tasks

GLM-5.2上线并开源：专注Coding与长程任务

Zhipu released GLM-5.2 under MIT license. It hit #1 among publicly available models on Code Arena, a front-end dev blind benchmark. Coding ability lands between Claude Opus 4.7 and 4.8. The model handles 1M-token lossless context and multi-day tasks. On FrontierSWE it trails Opus 4.8 by only 1%, beating GPT-5.5 and Opus 4.7; on Terminal-Bench 2.1 it's 4% behind Opus 4.8 but up 17.5% over GLM-5.1. A new thinking-budget control lets users dial reasoning depth. IndexShare architecture cuts unit FLOPs to 2.9×, and improved MTP layers boost acceptance length by 20%. It's already adapted to domestic hardware like Huawei Ascend, with the API live under the GLM Coding Plan.

Why it matters: Zhipu open-sourced GLM-5.2 under MIT, with Code Arena frontend blind test ranking #1 among public models, slotting between Claude Opus 4.7 and 4.8, and FrontierSWE only 1% behind Opus 4.8. 1M-token lossless context adds practical weight. Not scoring higher because only the hea...