GLM-5.3
Zhipu's 753B flagship invents a new trip-wire: pass $10B in revenue, then pass Z.AI's security review
What it is
Z.AI's flagship: a 753B-parameter MoE on the same base as GLM-5.2, with every gain — by the team's own account — coming from post-training. 1M-token context, a three-level reasoning_effort dial (low/high/max), and a benchmark story aimed at coding and cybersecurity: 66.9 DeepSWE, 84.5 CyberGym (claimed state-of-the-art), 105/130 on ExploitGym, and a fifty-percent jump over GLM-5.2 on Z.AI's internal code bench. Weights landed August 28 after a reported two-week safety hold.
Why it's interesting
The license is a genuinely new species and that's why it's shelved on Fine Print. The 'GLM-5.3 License' is MIT-shaped in its grant — use, modify, sell, deploy, fine-tune — but one clause changes the category: a Model-as-a-Service business whose aggregate revenue with affiliates exceeds US$10 billion over any consecutive 12 months 'must pass Z.AI's security review before using the Software or its derivative works for any commercial purpose.' Not a paid license like Qwen's, not a branding lock like Solar's — a security review, with scope and methodology 'reasonably determined' by Z.AI, no published criteria, no timeline, no appeal.
Use cases
- Frontier-class self-hosted reasoning and coding for orgs nowhere near the threshold — that is, almost everyone
- Cyber-defense research on the model claiming CyberGym state-of-the-art
- A case study in scale-triggered licensing — the newest variant of the genre
Who it's for
Infra-rich teams under the (very high) threshold; license watchers — this clause will be imitated
Setup
Advanced. 753B parameters means a serious multi-node deployment — coverage puts it at an eight-GPU minimum; BF16, FP8, and F32 tensors shipped
Limitations & cautions
The benchmark table is Z.AI's own with independent evals still landing; the security-review clause has no published scope, timeline, or appeals process, so what it demands is whatever Z.AI reasonably says it demands; and the license's silence on patents and termination leaves counsel more homework than a bespoke text usually does.
Editorial takeaway
The fine-print genre keeps innovating: first branding locks, then revenue triggers, now a discretionary security review. $10B is a threshold most readers will never see — but 'we reasonably determine' is a phrase every license lawyer will.