Zhipu has recently made an official announcement that the GLM-5.3 model has been open-sourced. Prior to this, the model had already rolled out its API service. However, owing to a substantial enhancement in its cybersecurity features, the release of its model weights was postponed. Now, following a thorough review process, the model weights have been released for public access.

The GLM-5.3 model builds upon the foundation of the GLM-5.2 base model. It employs a 743 billion-parameter Mixture of Experts (MoE) architecture, activating 39 billion parameters during the inference stage. It supports a 1 million-token context window, with model weights stored in the FP8 format and occupying approximately 756GB of space. The model is primarily focused on bolstering capabilities in complex programming, long-cycle agent tasks, and cybersecurity. Internal testing has demonstrated a performance improvement of around 50% compared to its predecessor.

The model is governed by a customized GLM-5.3 License, which permits free commercial use. Only Model-as-a-Service (MaaS) providers with annual revenues surpassing $10 billion are required to undergo a review process, while other users are not subject to any such restrictions.