Z.ai, the international brand of Chinese AI company Zhipu AI, said on June 17 it was launching and open-sourcing GLM-5.2 under the MIT License.

The model is designed for coding, long-document processing, and other multi-step tasks.

The company said GLM-5.2 supports a one million-token context window, up from 200K in GLM-5.1. It said the model ranked between Anthropic’s Claude Opus 4.7 and 4.8 on several evaluation.

The company said FrontierSWE, SWE-Marathon, and PostTrainBench, would make it the highest-ranking open-source model on those benchmark.

FrontierSWE is a public coding benchmark built around ultra-long-horizon technical tasks that give AI agents 20-hour budgets.

Z.ai also said GLM-5.2 includes structural and inference-side optimizations, including key-value cache and long-context operator improvements.

Z.ai has released GLM models under MIT since 2025, and the license allows use without field-of-use or geographic limits. The company also said its models run on more than 40 domestic chips, including Huawei Ascend and Cambricon.

🔗 Source: Z.ai

Recent Z.ai developments