Zhipu's GLM-5.3 adopts restrictive custom license as Chinese labs tighten open-model terms, amid a wave of new open model releases (Motif-3, dots3, Qwen3.8-Flash-Next, Hy4-preview, and more)
Zhipu released GLM-5.3 under a new custom license (replacing the MIT license used for GLM-5.2 and earlier) that requires companies with a Model-as-a-Service business and aggregate revenue over $10 billion across any 12 consecutive months to pass a Z.AI security review before commercial use of the software or derivatives. This continues a broader trend already seen with Kimi K3 and MiniMax M3 license restrictions. Separately, the article rounds up several other model releases in the same period: Motif-3 (MIT license, Motif-Technologies), dots3-note-prev (dots-studio/RedNote), Qwen3.8-Flash-Next preview (Qwen), GLM-5.3-Flash (initially released stealthily as 'Ox-Alpha' on OpenRouter/OpenCode before being revealed as Zhipu's), Hy4-preview (Tencent), NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 update, Ling-3.0-flash (inclusionAI), and Qwen3.8-2.4T-A95B (Qwen's largest openly released model to date, under a custom license).
Entities: Zhipu, Z.AI, GLM-5.3, GLM-5.3-Flash, GLM-5.2, Kimi K3
0 primary
What happened
Zhipu released GLM-5.3 under a new custom licence, replacing the permissive MIT licence used for GLM-5.2. The new terms require any company running a Model-as-a-Service business with over $10 billion in aggregate revenue across a 12-month period to pass a Z.AI security review before commercial use of the software or derivatives. This follows similar restrictive clauses already added by Kimi K3 and MiniMax M3. The same period also saw a cluster of other open model releases from Motif-Technologies, dots-studio, Qwen, Tencent, NVIDIA and inclusionAI, plus a brief hype cycle around a "stealth" model later revealed as Zhipu's GLM-5.3-Flash.
Why it matters
This is now the third major Chinese lab (after Kimi and MiniMax) to add commercial-use gating to an open-weight model, suggesting a pattern rather than a one-off, while Western labs like Google and Meta continue favouring permissive Apache 2.0 terms. In practice this only directly affects a small number of very large MaaS providers with over $10 billion in revenue, so most developers, startups and researchers can keep using GLM-5.3 exactly as before. The real consequence is for procurement and legal teams at large cloud and AI platform companies, who now need to track licence terms model-by-model rather than assuming open-weight releases are uniformly permissive.
What is noise
Framing this as a dramatic "closing" of Chinese open models is overstated: the $10 billion revenue threshold excludes nearly every company that will ever touch this model, and this is the third instance of the same trend, not a new discovery. Bundling nine unrelated model releases and a stealth-model hype anecdote into one story dilutes the actual news, which is narrowly a licensing change affecting a handful of hyperscale-scale MaaS operators.
Watch next
- 01Whether any other major Chinese lab (e.g. DeepSeek, Alibaba/Qwen) adopts similar revenue-gated licensing terms in their next flagship release
- 02Whether any named large MaaS provider (e.g. a company plausibly over $10B revenue) publicly confirms undergoing or being blocked by a Z.AI security review
- 03How legal analysts interpret the undefined 'affiliates' term in the English licence versus the more precise Chinese legal term, and whether Zhipu clarifies or amends the wording
Coverage
1 storyMore distribution signals
Full feed →- OpenAI ends its contract supplying models to Cursor (Anysphere) after Cursor's acquisition by SpaceX28 Aug 202678
- Google offers Gemini's personalized AI image generation for free to US users29 Jun 202677
- OpenAI launches GPT-6 Astra, claims major capability jump and declares "AGI era"3 Sept 202676
- OpenAI ends its API partnership with Cursor after SpaceX's $60B acquisition of the coding startup3 Sept 202676