On August 14, Zhipu officially released the large model GLM-5.3. The base parameter count remains over 74 billion, the same as GLM-5.2, and the rumored upgrade to 1 trillion parameters did not appear—this card might be reserved for GLM-5.5. However, Zhipu's post-training approach significantly improved the performance of GLM-5.3 by 50% compared to the previous generation, achieving the highest rankings among current open-source models in multiple mainstream benchmark tests. Its programming and agent capabilities are close to Claude Fable5, and its programming experience exceeds other domestic models.

Several key benchmark scores have risen significantly, which is very persuasive. On Terminal-Bench3.0, which measures a model's ability to complete complex tasks in real terminal environments, the score jumped from 4.6 to 28.3; on DeepSWE v1.1, which focuses on long-range software engineering and continuous code modification capabilities, it rose from 46.2 to 66.9; on Agents' Last Exam, which covers various real professional scenarios and emphasizes cross-tool collaboration and long-range tasks, it increased from 23.8 to 28.5. In GDPval-AA v2, which covers 44 professions and examines high-value knowledge work, GLM-5.3 scored 1769 points, demonstrating its professional task execution capability based on programming skills. Overall, it has made significant progress in complex software engineering, terminal operations, and broader real-world agent tasks.

image.png

The most telling evidence is Zhipu's self-developed Z.ai Code Bench—a set of evaluations that place the model in a real local development environment, executing end-to-end tasks under different thinking modes, as realistically as possible simulating the experience when developers use a Coding Agent. Test results show that GLM-5.3 has found a better balance between effectiveness and Token utilization: at the High mode, the accuracy rate reached 31.4%, exceeding the maximum mode of Claude Opus4.8 by 29.5%, while each task only outputs about 50,000 tokens on average, whereas Opus4.8 requires about 120,000 tokens—indicating that GLM-5.3 can accomplish the same tasks with a shorter execution path.

image.png

In terms of product implementation, GLM-5.3 is now available on Zhipu's official programming tool ZCode and efficiency tool AutoClaw, and is also open to all users of GLM Coding Plan and subscription users. Third-party coding platforms such as TraeWork, TraeCode, Kousi, WorkBuddy, CodeBuddy, Qoder, QwenWork, CatPaw, JoyCode, and OpenCode have also opened up early access. The API will be launched soon, and the complete model weights will be open-sourced after necessary security reinforcement within two weeks. Zhipu's strategy is to limit the potential attack capability of the model while retaining its defensive value. At 13:00 today, the quota for all users of GLM Coding Plan has been reset, and all users can see their quotas restored in the backend usage statistics.