Z.ai Releases GLM-5.3 to Boost Agentic Coding by 50%

Z.ai has officially released the open weights for GLM-5.3, its latest mixture-of-experts model designed for agentic coding and cyber defense. Built on the same 743-billion-parameter base checkpoint as GLM-5.2, the new model utilizes roughly 40 billion active parameters per token. Remarkably, the developer achieved a 50 percent boost on its in-house Z.ai Code Bench without any additional pretraining tokens, demonstrating that sophisticated post-training can unlock massive performance gains from existing base weights.

The model establishes new open-source benchmarks, particularly in long-horizon tasks and security. On Terminal Bench 3.0, GLM-5.3 jumped to a score of 28.3 from GLM-5.2's 4.6, while trailing the closed-source leader, GPT-5.6 Sol, at 34.6. On DeepSWE v1.1, it rose to 66.9 from 46.2, compared to GPT-5.6 Sol at 72.7. It also tied for the lead on AutomationBench v1.0.6 with a score of 48.2, up from 26.2. In cybersecurity, GLM-5.3 leads CyberGym with an 84.5 percent Pass@1 rate across 1,507 vulnerability discovery tasks. It also doubled its predecessor's performance on exploitation benchmarks, scoring 54.4 on ExploitBench, though it remains behind Claude Fable 5's score of 78.0.

For practitioners, GLM-5.3 offers a massive 1-million-token context window, making it suitable for repository-scale refactoring. The model is available on Hugging Face with BF16, FP8 E4M3, and F32 tensors, and it integrates with popular frameworks including vLLM, SGLang, Transformers, Unsloth, TokenSpeed, and KTransformers. It also supports Ascend NPUs via vLLM-Ascend and xLLM.

Developers integrating the model should note key API changes. Z.ai has replaced the previous thinking-off toggle with a new reasoning_effort parameter that accepts low, high, or max values, defaulting to max. Disabling the thinking process entirely is no longer supported, and pipelines attempting to do so will fail. Additionally, the clear_thinking parameter defaults to false, though Z.ai recommends setting it to true for interactive chat applications.