Z.ai Open-Sources GLM-5.3 Weights for Coding and Autonomous AI Agents

Chinese lab Z.ai has published the full weights for its flagship GLM-5.3 model on Hugging Face. The neural network targets complex coding tasks, terminal automation, and autonomous AI agent control, delivering performance that matches leading proprietary commercial systems.

Full 756GB Release and Supported Frameworks

Unlike the previously introduced lightweight multimodal GLM-5.3-Flash (a 320-billion parameter MoE architecture with 18 billion active parameters), the developers have released the full-scale flagship model itself.

According to the Hugging Face repository, the total file size of the weights in Safetensors format is approximately 756GB. The model is optimized for local and server deployment using key open-source engines:

• vLLM and SGLang for high-speed inference;
• Transformers and KTransformers;
• Unsloth and TokenSpeed frameworks;
• Specialized stacks for Huawei Ascend NPU hardware accelerators.

Leap in Coding and Agent Benchmarks

While the base architecture remains unchanged from GLM-5.2, Z.ai engineers applied an extensive post-training phase focusing on long-horizon reasoning and executing tasks within sandboxed environments.

Data from the official Z.ai blog shows the model pulled significantly ahead of its predecessor across standardized tests:

• Terminal Bench 2.1: scores climbed to 88.2 points compared to 81.0 for GLM-5.2 (for reference, proprietary GPT-5.6 Sol scored 88.8);
• Terminal Bench 3.0: jumped from 4.6 to 28.3 points;
• DeepSWE v1.1: rose from 46.2 to 66.9 points;
• AutomationBench: improved from 26.2 to 48.2 points.

In Z.ai’s internal synthetic code generation benchmark, Z.ai Code Bench, accuracy increased by an average of 50%.

Unexpected Cybersecurity Skills and Release Delay

GLM-5.3 was initially announced on August 14, but the release of the weights was pushed back by two weeks. The delay was caused by an unexpected surge in the model’s vulnerability analysis capabilities.

In CyberGym benchmarks, the model scored 84.5 points, and in ExploitBench it reached 54.4 (up from 24.4 in version 5.2), doubling its performance in exploit scenarios. Developers used the extra time to test guardrails and conduct safety alignment to reduce the risks of misuse.

We use GLM-5.3-Flash build a dream kitchen. A 3D world built in Blender. This is not a generated video. pic.twitter.com/JILdFOY21l

— Lou (@louszbd) August 27, 2026

Licensing Terms

The model weights are distributed under an open-weights model, allowing developers and organizations to freely download and run them on their own hardware. However, the project is governed by a proprietary GLM-5.3 license rather than traditional permissive licenses such as MIT or Apache 2.0. For most local enthusiasts, the Flash version remains the more practical choice, while the full 756GB model is aimed at enterprise server clusters.