On June 13, Zhipu AI fully opened its latest flagship model, GLM-5.2, to all GLM Coding Plan users (covering Lite, Pro, Max, and Team editions). The company positions it as “the most capable open-source model from Zhipu AI to date,” which adopts the same MoE architecture as GLM-5.1. The most significant upgrade this time is the first-ever realization of a “truly usable” 1 million token context window — the company explicitly distinguishes it from competing products that nominally support such a window but suffer from severe performance degradation in practice. It is designed to stably handle engineering-level tasks such as understanding complete code repositories, multi-file collaborative editing, and long-range debugging. The model will be officially open-sourced within this week under the MIT license, with APIs going live simultaneously. On the same day, Zhipu AI also released the ZCode 3.0 programming tool, which fully transitions to the self-developed ZCode Agent kernel and is deeply adapted for GLM-5.2, while no longer being compatible with third-party Agent frameworks. GLM Coding Plan users enjoy a 150% exclusive quota bonus, and new users can claim a 5-day free trial of ZCode (5 million tokens per day).
In the latest updated blind evaluation leaderboard from the third-party testing organization Arena, GLM-5.2 (Max) jumped to 10th place, becoming one of the open-source models currently in the Arena global top 10. The previous generation, GLM-5.1, ranked 13th, and DeepSeek-V4 Pro ranked 14th — Chinese open-source models have formed a major cluster on the Arena leaderboard, with the only remaining overseas open-source contenders being NVIDIA Nemotron 3 Ultra and Google Gemma 4, both at the bottom. Notably, Zhipu AI explicitly stated in its announcement that it chose to respond with an open-source route “at a time when some cutting-edge models have suddenly become unavailable,” which is widely interpreted as a strategic statement against the backdrop of export controls on Anthropic Fable/Mythos.
z.ai | Hugging Face | Arena