GLM-5.3 released with open weights: Z.ai's strongest agent model for programming and cyber defense

Z.ai, a Chinese AI firm, has released GLM-5.3 as an open-weight model, centering the announcement on two specific domains: agent programming and cyber defense. The model is available for download, local execution and fine-tuning, without license restrictions that limit commercial use. Unlike earlier versions that highlighted general capabilities, this release is explicitly aimed at end-to-end agent tasks in real-world environments.

same base, different post-training

The underlying architecture is identical to GLM-5.2; all improvements stem from post-training. According to data published by the company, the new model shows a 50-point percentage gain over its predecessor on the internal Z.ai Code Bench. On public benchmarks it achieves open-source state-of-the-art (SOTA) results on two core tests: Terminal Bench 3.0 and Agents' Last Exam (ALE). Both tests simulate an agent operating in a real CLI environment with long context and external tools.

cyber capability grew faster than expected

The most notable surprise lies in cyber security. As post-training expanded, the model’s vulnerability-discovery ability progressed faster than forecasts. GLM-5.3 attains SOTA on CyberGym, a benchmark that evaluates vulnerability discovery and exploitation, with the largest margin appearing in high-exploitation stages where the model more than doubles GLM-5.2’s performance. The implication is that the model not only identifies bugs but also knows how to exploit them in a controlled setting.

broad deployment support and hardware acceleration

GLM-5.3 supports a long list of inference frameworks: SGLang, vLLM, TokenSpeed, Transformers, KTransformers and Unsloth, each with a dedicated guide. For Huawei’s Ascend NPU there is dedicated support via vLLM-Ascend, xLLM and SGLang. At the parameter level, GLM-5.3 offers control over the reasoning budget through the reasoning_effort setting with three levels (low, high, max, default max) and a clear_thinking flag that defaults to false in chat and requires explicit activation.

detailed and transparent evaluation methodology

Z.ai publishes the full set of evaluation parameters for every benchmark, including temperature, top_p, maximum context length, timeout, number of rollouts, and scoring method. For example, Terminal Bench 3.0 was evaluated with avg@3 over three rollouts per task, in an isolated container, with a maximum of 600 agent turns, a 10-hour timeout, and no search tools. ALE was run on 105 tasks in Docker with a 4–8 hour timeout. CyberGym was evaluated in Claude Code 2.1.207 with maximum reasoning effort and no network tools. This granularity enables complete reproducibility, a rarity in announcements of this type.