Chinese AI laboratory Z.ai has released the open weights for its flagship GLM-5.3 model on Hugging Face, introducing a tiered licensing structure that places specific restrictions on commercial hyperscalers. The release follows a two-week safety evaluation period that began after the model's initial API launch earlier this month.
While previous models in the family, including GLM-5.2 and GLM-5.3-Flash, were distributed under the permissive MIT license, the flagship release introduces the custom GLM-5.3 License. Under these terms, organizations hosting the model directly that generate more than $10 billion in aggregate revenue over any consecutive 12-month period must pass a formal security review conducted by Z.ai before using the software or derivative works for commercial purposes.
Developers, researchers, and hosting aggregators such as OpenRouter remain unaffected by the revenue threshold, maintaining rights to run, fine-tune, and deploy the weights.
Architecture and Resource Profiles
GLM-5.3 utilizes the same 753-billion-parameter Mixture-of-Experts (MoE) base architecture as GLM-5.2, paired with a 1 million-token context window and a maximum output limit of 128,000 tokens. All performance gains between the two releases stem from post-training scaling across RL environments and long-horizon tasks.
The model weights are available in BF16 and FP8 precision, with immediate runtime support implemented across vLLM, SGLang, KTransformers, and Hugging Face Transformers. Local hosting hardware requirements vary significantly by quantization level:
- Unsloth 2-bit quantization: Requires 245GB of unified memory, enabling single-node deployment on 256GB unified memory hardware while retaining roughly 86 percent top-1 accuracy.
- 8-bit quantization: Requires 810GB of VRAM across distributed accelerator clusters.
- BF16 base precision: Requires multi-node enterprise GPU clusters for full uncompressed inference.
On the commercial API side, Z.ai prices standalone inference at $1.40 per million input tokens and $4.40 per million output tokens, positioning it above the lightweight GLM-5.3-Flash variant ($0.15 input / $0.47 output) while undercutting competing closed-weights frontier APIs.
Post-Training and Cybersecurity Benchmarks
Z.ai attributed the model's reasoning improvements to its post-training stack, which combines IndexShare for long-context efficiency, SAO reinforcement learning for multi-step agent trajectories, and the asynchronous slime framework.
On software engineering and cybersecurity evaluations:
- CyberGym: GLM-5.3 achieved 84.5 percent in white-box vulnerability detection, compared to 77.2 percent for GLM-5.2.
- ExploitBench: Scored 54.4 percent, more than doubling GLM-5.2's 24.4 percent mark.
- Terminal Bench 3.0: Advanced from 4.6 on GLM-5.2 to 28.3 on GLM-5.3.
- DeepSWE v1.1: Rose from 46.2 percent to 66.9 percent task resolution.
Z.ai reported that the model identified 2,436 security vulnerabilities across 269 open-source repositories during red-teaming exercises, with coordinated disclosures tracked on its public security ledger.
Shifting Licensing Dynamics in Open Weights
The introduction of the $10 billion revenue clause reflects a growing divergence in open-weight licensing strategies among frontier labs. While DeepSeek continues to publish flagship models under standard MIT licenses, competitors like Moonshot require prominent UI attribution for providers exceeding 100 million active users or $20 million in monthly revenue.
By retaining MIT licensing for GLM-5.3-Flash while gating the flagship 753B model, Z.ai is establishing a bifurcated distribution model: commodity flash models remain unencumbered, while frontier-tier agentic weights require explicit commercial clearances from top-tier cloud operators.
