Z.ai's GLM-5.3 reported to carry a custom restrictive license, not MIT as earlier described
2026-09-08
Nathan Lambert's Interconnects reports that Zhipu/Z.ai's GLM-5.3 switched from the MIT license used for GLM-5.2 and earlier to a custom license, quoting a clause requiring model-as-a-service providers whose aggregate revenue (including undefined "affiliates") exceeds $10 billion over any 12 months to pass a Z.AI security review before commercial use. Lambert notes the license text exists in both English and Chinese versions, with the Chinese term for "affiliated parties" carrying a defined meaning in Chinese law that the English text lacks. This contradicts earlier reports describing GLM-5.3-Flash as MIT-licensed.
Significance 3: A credible license-text change reported by a researcher-lens curator quoting the clause directly; a meaningful strategy signal for the open-weights ecosystem but below verdict-changing threshold, and uncorroborated by a primary Z.ai source.
Implications · machine-drafted, not owner judgment
Z.ai's shift from MIT to a custom license with a revenue-gated security-review clause signals that even the most open Chinese labs are hedging against unconstrained commercial use of their weights by the largest incumbents, narrowing what "open weights" practically means even as the open/closed capability gap narrows. With "affiliates" left undefined and the license existing in two language versions, enterprises building on GLM-5.3 face real legal ambiguity before ever approaching the $10B threshold. Per Lambert's framing, this nuances what "open" means going forward: capability parity is arriving, but license terms are trending toward more friction for the largest downstream users, not less.
- Other Chinese labs (Kimi, MiniMax, Qwen) tightening license terms similarly
- A public clarification of the license's "affiliates" definition
- Inference providers publicly changing GLM-5.3 hosting terms in response
Sources
- curator Latest open artifacts (#24): Motif-3, GLM-5.3, Hy4-preview and open model licenses retrieved 2026-09-08