Tier 2 Restricted open-weight
What the publisher claims: open-weight
The one-liner: Open-weight frontier coding model, not a full open-source AI release under the strict rubric.
The six questions
GLM-5.3 weights were published on Hugging Face on 28 August 2026 after a two-week safety evaluation, and GLM-5.3 Flash shipped with MIT weights at launch.
GLM-5.3 ships under a custom license that grants broad use, modification, and commercial rights but requires a Z.AI security review, with scope set by Z.AI, for Model-as-a-Service operators above $10B aggregate revenue; because the review gate sits with the vendor, the grant is conditional rather than clean.
Independently hosted on widely available hardware with vLLM and SGLang support and a large provider ecosystem, with quantized community builds available.
Architecture, parameter counts, post-training changes, and evaluation results including CyberGym scores are published, and the release was held back for documented safety evaluation, but no training data or reproducible recipe accompanies the weights.
Limited: an equivalent model can be approximated from the published post-training description but not rebuilt.
Z.AI publishes GLM weights and post-training change notes on Hugging Face and takes some outside tooling pull requests, but the release train and training data remain proprietary and internal.
License
Custom GLM-5.3 License for GLM-5.3; MIT retained for GLM-5.3 Flash and earlier GLM-5.x releases — license source ↗
- Model-as-a-Service operators with more than $10B aggregate 12-month revenue must pass a Z.AI security review whose scope and method Z.AI determines
- Release-specific terms; the license must be read rather than inferred from open branding
What is actually released
- weights (GLM-5.3 753B, GLM-5.3 Flash 320B multimodal, GLM-5.2)
- architecture and parameter documentation
- post-training change notes
- evaluation results including cyber benchmarks
Why it matters
Z.AI explicitly uses the open-weight label and discusses vulnerability-discovery training, while the release provides vendor-controlled security review and no OLMo-level reproducibility - the label and the evidence point in different directions.
Sources
- https://huggingface.co/zai-org/GLM-5.3/blob/main/LICENSE
- https://z.ai/blog/glm-5.3
- https://docs.api.nvidia.com/nim/reference/z-ai-glm-5-3
- https://geotoolbox.ai/blog/what-is-glm-5-3
- https://www.nist.gov/news-events/news/2026/09/caisis-assessment-zais-glm-53-cyber-capabilities
Derived from the openness dataset behind But Is It Open? — snapshot 2026-11-07 ↗. Scores are qualitative readings of the cited evidence, not verdicts; tier measures openness, not intelligence.