GLM-5.3 weights land with a hyperscaler clause

Z.ai published the weights for GLM-5.3 on Hugging Face on Friday under a new license that requires any company hosting the model commercially with more than $10 billion in aggregate revenue over any 12 consecutive months to pass a Z.AI security review. Its predecessor, GLM-5.2, shipped under the permissive MIT license, The New Stack reported.

At a glance

• The license draws a line between hosting and routing: marketplaces that only route requests, as OpenRouter does, fall outside it, while providers that serve the weights themselves and clear the revenue bar do not.
• Serving GLM-5.3 costs $1.40 per million input tokens and $4.40 per million output, while the Flash variant runs at $0.15 and $0.47, and 2-bit quants still need 245GB of memory.
• Individual users keep the rights to run, deploy and fine-tune, and GLM-5.3-Flash remains under MIT, so the review requirement applies only to the largest commercial hosting providers.

The open-weight tier has largely run on permissive terms with no revenue tests attached, which is what made Chinese releases easy for anyone to serve. Z.ai kept MIT for Flash and applied the new conditions only to its strongest model, which reads as a claim on the inference layer as much as a safety measure. The timing sharpens it: Nvidia is moving on Hugging Face and Stripe on OpenRouter, so the download and rental rails are consolidating in American hands while the models increasingly come out of Chinese labs.

Companies above $10 billion in revenue must clear a security review before hosting

The GLM-5.3 License applies its revenue test to companies that host the model rather than merely route it. Those above $10 billion in aggregate revenue over any 12 consecutive months "must pass Z.AI's security review before using the Software or its derivative works for any commercial purpose."

For individual users the terms cover running, deploying and fine-tuning, and GLM-5.3-Flash, the model that went viral as ox-alpha, remains under MIT. Custom licensing is not new here: ChatGLM3-6B shipped under terms requiring registration for commercial use, and every Z.ai model after it went out under MIT.

Other Chinese labs are less demanding. Moonshot requires a model-as-a-service provider offering Kimi K3 with more than 100 million active users or more than $20 million in monthly revenue to display "Kimi K3" prominently in its interface, and DeepSeek still ships its flagship models under MIT.

Z.ai reports 84.5 percent on CyberGym and 2,436 vulnerabilities found

Z.ai held the weights back for two weeks after the API launch for safety evaluation and hardening, a break from GLM-5.2, whose weights were available on launch day. The company reports 84.5 percent on CyberGym, a vulnerability discovery benchmark, and calls it the best published result.

That figure is self-reported and no one outside the company has reproduced it. Z.ai also says it used the model to find 2,436 vulnerabilities across 269 open-source projects, including the Linux kernel, though only a few dozen of those findings can be inspected publicly so far.

Despite the safety framing, the license itself contains no acceptable-use section and says nothing about cyber capability or offensive security. Announcing the release in a post on X, Zixuan Li of Z.ai tied the delay directly to the model's cybersecurity capabilities.

Given the model's advanced cybersecurity capabilities, we conducted two additional weeks of comprehensive safety evaluations before releasing the weights.

Even 2-bit quants of GLM-5.3 need 245GB of memory

The architecture is unchanged from GLM-5.2: 753 billion parameters in a mixture-of-experts layout, a 1 million-token context window and a maximum output of 128,000 tokens. The weights ship in BF16 and FP8 and run on vLLM, SGLang, KTransformers and Hugging Face's Transformers library.

Local deployment demands substantial hardware. Unsloth's 2-bit quants, which the group says still reach about 86 percent top-1 accuracy, require 245GB of memory, just fitting a Mac with 256GB of unified memory, while the 8-bit quants need 810GB.

Several third-party inference services already host GLM-5.3 and expose it through marketplaces such as OpenRouter. Nvidia moved this week to acquire Hugging Face for a reported $12.9 billion, and Stripe agreed to acquire OpenRouter; if both deals close, the main repository and the main rental venue for open-weight models will be American-owned.

Whether GLM-5.4 keeps MIT Z.ai has not said whether the terms will carry over to its next flagship model, and no date has been given for one. Flash stays on MIT for now, so the company has not abandoned permissive licensing outright, only stopped applying it to its strongest model. Hosting providers below the $10 billion threshold are unaffected by the review requirement.