Chinese AI startup Zhipu (Z.ai) has officially open-sourced the weights of its flagship large language model GLM-5.2 and launched an API service. Early last week, GLM-5.2 was already made available to all GLM Coding Plan subscribers, but this update marks the public release of the model weights and general API access.
https://twitter.com/Zai_org/status/2066938937344495629
As a mixture-of-experts (MoE) model, GLM-5.2 packs a total of 753 billion parameters, though each inference activates only about 40 billion parameters. To handle different tasks, Zhipu offers two thinking modes: High and Max. The High mode balances response speed and answer quality, while the Max mode goes deep into chain-of-thought reasoning — especially recommended for tasks like automated coding and complex debugging.
Beyond architecture, GLM-5.2 also supports a 1-million-token long context, which has been specifically trained for stability on long-range complex tasks. To avoid slowdowns when processing lengthy texts, Zhipu introduced a technique called IndexShare under the hood. By having sparse attention layers share the same retrieval index every four layers, the single-token computation cost at 1 million tokens is reduced by 2.9 times — speeding up long-text responses while cutting resource consumption.
The model weights are released under the MIT license, allowing free commercial use and modification. Zhipu's official API is now live with pricing that matches the previous-generation GLM-5.1. GLM-5.2 has also been integrated into the global API platform OpenRouter, making it accessible to overseas developers.