GLM-5.3 weights land with a hyperscaler clause

Z.ai published the GLM-5.3 weights on Hugging Face on August 28 with one main condition attached. Any company that hosts the model commercially and has more than $10 billion in aggregate revenue over any 12 consecutive months must first pass a security review by Z.ai. Everyone below that line can run, deploy, fine-tune and sell services on the model.
The threshold is part of a new license that Z.ai calls the GLM-5.3 License. It replaces the MIT license that GLM-5.2 used in June. The New Stack reports that the clause covers companies that host the model. Companies that only route requests to it, as OpenRouter does, or that build it into a product are outside it. Several third-party inference services already serve GLM-5.3, and OpenRouter lists it.
Who has to pass the review
The key sentence of the license says that a host above the revenue line “must pass Z.AI’s security review before using the Software or its derivative works for any commercial purpose.” In practice that means the hyperscalers. Neoclouds, the smaller GPU-rental clouds, fall under the same rule once their revenue crosses $10 billion. For individual users, The New Stack says, nothing really changes. The new text is written more specifically for models and spells out the rights to run, deploy and fine-tune.
Other Chinese labs set looser terms for their flagship models. Moonshot’s license for Kimi K3 is triggered when a model-as-a-service provider has more than 100 million active users or more than $20 million in monthly revenue. Even then, the only requirement is that “‘Kimi K3’ must be prominently displayed on the user interface of such product or service.” DeepSeek still uses the plain MIT license.
Z.ai has used custom licenses before. In 2023 and 2024, models such as ChatGLM3-6B came under a license that required registration for commercial use. After that, every new Z.ai model shipped under MIT, including the GLM-4.5, GLM-4.7, GLM-5, GLM-5.1 and GLM-5.2 checkpoints, according to Wikipedia and Systems analysis Wiki.
Z.ai presents the change as a safety measure. The New Stack points out that the license has no acceptable-use section and says nothing about cyber or offensive security. The outlet also asks whether the motive is security or monetization. With GLM-5.3-Flash, Z.ai stressed that inference ran on Chinese chips, which suggests the company wants to own the inference layer for its models. As open-weight models get closer to what U.S. frontier labs ship, that layer is likely worth more.
Two extra weeks before the weights
Z.ai says it held the weights for two weeks after the API launch to run safety evaluation and hardening. That is new for the company: GLM-5.2’s weights were available on launch day.
Given the model’s advanced cybersecurity capabilities, we conducted two additional weeks of comprehensive safety evaluations before releasing the weights.
Zixuan Li wrote this in an August 28 post on X. The headline cyber figure is 84.5 percent on CyberGym, a benchmark for finding vulnerabilities. Z.ai calls it the best published result. The number is self-reported and nobody outside the company has reproduced it yet. On the same benchmark, Microsoft’s MDASH multi-agent scaffold posted 88.45% in August. MDASH coordinates several agents, while 84.5 percent is the score of a single model.
Z.ai also says it used GLM-5.3 to find 2,436 vulnerabilities across 269 open-source projects, the Linux kernel among them. Its release notes say the work was done with several cybersecurity teams and that 1,097 of the findings, under half, are medium or high severity. Only a few dozen can be inspected publicly so far. The same notes say GLM-5.3 matches Mythos 5 at white-box code review and vulnerability discovery.
Z.ai focused this release on post-training, the tuning done after the base model is trained. The release notes claim a 50% gain over GLM-5.2 on Z.ai Code Bench and the top score among open-source models on public benchmarks including Terminal Bench 3.0. The New Stack describes a model that often beats other Chinese open-weight models and keeps pace with current models from the large U.S. labs.
What it takes to run 753 billion parameters
The architecture has not changed since GLM-5.2. It is a 753-billion-parameter mixture-of-experts model, a design that sends each token through only a few of its expert sub-networks. The context window is 1 million tokens and output goes up to 128,000 tokens. The weights come in BF16 and FP8 and run on vLLM, SGLang, KTransformers and Hugging Face’s Transformers library.
Running it locally is still a job for large hardware. Unsloth’s 2-bit quants need 245GB of memory, and Unsloth says they keep about 86 percent top-1 accuracy. That just fits on a Mac with 256GB of unified memory. The 8-bit quants need 810GB.
Most users will rent it. Through the API, GLM-5.3 costs $1.40 per million input tokens and $4.40 per million output tokens, cheaper than nearly all of its closest competitors. GLM-5.3-Flash costs $0.15 and $0.47 for the same volumes.
Who owns the distribution
The license change came in the same week that Nvidia moved to buy Hugging Face, for a reported $12.9 billion, and Stripe agreed to buy OpenRouter. If both deals close, American companies will own the main repository for downloading open-weight models and the main marketplace for renting them, while more and more of the models come from Chinese labs.
The New Stack’s reading is that if GLM-5.4 ships under the same terms, Z.ai’s MIT years will look like a customer-acquisition phase. Open weights would then start to look less like a gift to the ecosystem and more like a distribution channel with terms attached.
For now, Z.ai is splitting its lineup. GLM-5.3-Flash, the smaller model that ran on OpenRouter under the name Ox Alpha, is still under MIT. The GLM-5.3 License applies only to the flagship.
Up next
Cyber access gate keyed to a 1996 US export list
A security researcher in Georgia picked his country in OpenAI’s cyber verification flow and got “Unable to verify”. His audit of all 250 entries found 73 that accept no government ID, 44 of them official ChatGPT markets.
Related stories
- Xiaomi’s new MiMo models carry an unverified top-6 claim
- Open-weight models win the tokens, Anthropic keeps the cash
- Zhipu AI will release GLM-5.3 weights in stages after safety reviews
- One model cleared all five test categories at 100%
- Kimi K3.1 Pro weights land on HuggingFace at 2.8T params
- Tencent ships Hy4 at Hy4 preview’s API price
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

