GLM-5.2 Model by Zhipu AI is Officially Open Sourced: Focuses on 1M Lossless Context and Long-Range Coding Tasks

 On June 17, Zhipu AI officially launched and open-sourced its new large model GLM-5.2, which focuses on Coding (code generation) and long-term task execution. On the cutting-edge front-end development evaluation system Code Arena, GLM-5.2 ranks second globally and first among open-source models.

Since early 2025, Zhipu has focused on coding foundations, successively launching GLM-4.5 and GLM-4.7. The recently released GLM-5.2 further extends its technical capabilities to the field of long-term tasks, aiming to solve complex engineering execution challenges that span days or even months.

In terms of technology, GLM-5.2 achieves a 1M lossless context, effectively solving the problem of long-text degradation by expanding the training environment of the Coding Agent. In long-term task benchmarks such as FrontierSWE, its performance is between Claude Opus 4.7 and 4.8.

In practical engineering applications, the model can process up to 880,000 tokens in one go, and independently complete the full lifecycle development of multi-platform applications including Web, mobile, and mini programs. Additionally, the model introduces an effort level control mechanism, significantly improving performance in mainstream programming evaluation benchmarks such as Terminal-Bench 2.1 compared to its predecessors.

In terms of underlying architecture and ecosystem compatibility, GLM-5.2 uses the IndexShare architecture to reduce the unit token FLOPs to 2.9 times under 1M context, and improves the MTP layer to enhance speculative decoding efficiency. Currently, the model is open-sourced under the MIT open-source license on Hugging Face and ModelScope, and it completed inference adaptation with domestic computing platforms such as Huawei Ascend, Pingtouge, and Moerthread on the day of release.

Moving from "intelligent assistant" to "digital employee," the release of GLM-5.2 marks the evolution of large models toward fully autonomous intelligent agent systems (Autonomous Agent Systems), laying a solid foundation for a comprehensive, self-driven digital productivity transformation.