Z.ai’s GLM-5.3 goes open weight, but its new license aims at hyperscalers
The New Stack Frederic Lardinois ● Covered by 2 sources
Z.ai put GLM-5.3’s weights online, but only big cloud firms face new license hoops. It’s open weight, yet the fine print looks aimed squarely at hyperscalers.
Based on reporting by The New Stack, Frederic Lardinois — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Z.ai has released the weights for GLM-5.3, its flagship model that surfaced in August after first appearing as the viral ox-alpha model and then turning out to be GLM-5.3-Flash. The model is now on Hugging Face and already shows up through several third-party inference services, including OpenRouter.
The bigger change is the license. GLM-5.2 used MIT. GLM-5.3 does not. Z.ai’s new license allows individuals to run, deploy, fine-tune and use the model commercially, but companies that want to host it themselves and have more than $10 billion in aggregate revenue over any 12 consecutive months must clear a Z.AI security review first. That does not apply to simple routing or embedding, and it is the kind of clause hyperscalers tend to notice immediately.
Z.ai says it held the open weights back for two weeks for safety evaluation and hardening. The company also says GLM-5.3 scored 84.5 percent on CyberGym, which it calls the best published result, and says it used the model to find 2,436 vulnerabilities across 269 open-source projects, including the Linux kernel. Only a small number of those findings are publicly inspectable so far, and the benchmark number has not been reproduced outside the company.
Under the hood, GLM-5.3 keeps the same 753-billion-parameter mixture-of-experts setup as GLM-5.2, with a 1 million-token context window and a maximum output of 128,000 tokens. The weights ship in BF16 and FP8, and they run on vLLM, SGLang, KTransformers and Hugging Face’s Transformers library. Z.ai says the model is available for download, local deployment, fine-tuning and commercial use.
This is still not a model most people will run on their own machines. Even the 2-bit quantizations need 245GB of memory, which Unsloth says can fit on a Mac with 256GB of unified memory. The 8-bit versions need 810GB. On API pricing, GLM-5.3 sits at $1.40 per million input tokens and $4.40 per million output tokens, while GLM-5.3-Flash is much cheaper at $0.15 and $0.47.
My take — AI-written commentary, not fact-checked reporting
This is the sort of “open” that actually matters in 2026: the weights are out, but the leverage sits in the license. Z.ai is making a clean bet that the real money is no longer in sharing the model, but in deciding who gets to run it at scale. That’s a very Chinese-lab-meets-hyperscaler way to do open source, and it’s not even pretending to be generous anymore.
Read more about this at: The New Stack