Particle.news

ZhipuAI Open-Sources GLM-5.2 With Stable 1M-Token Context

The release signals an open-source rival to proprietary models by offering long-context capability and immediate support on Chinese accelerator platforms.

Overview

  • ZhipuAI published GLM-5.2 and released model files and code on GitHub, Hugging Face and ModelScope, a move the company says makes the model broadly available to developers.
  • ZhipuAI, which announced GLM-5.2 on Wednesday, June 17, 2026, reported that the model placed first among globally available models in Code Arena's blind test for coding tasks.
  • The company presents GLM-5.2 as engineered for long-range tasks with a stable 1 million token context window so it can handle very large prompts, long documents and multi-step agent workflows.
  • ZhipuAI says it cut per-token compute to 2.9x at 1M context and completed Day-0 inference adaptations for a range of domestic accelerators including Huawei Ascend, Phytium, Moore Threads, Cambricon, Kunlunxin, Muxi, Hygon and Biren.
  • ZhipuAI expects Huawei’s upcoming Ascend 950 supernodes to be a strong future compute base for GLM-5.2, but that forecast is prospective and the company’s benchmarking claims have not been independently verified.