The licenses of open AI models have become a subject in their own right. While Western players are adopting permissive licenses, several leading Chinese publishers are tightening their terms, as seen with Zhipu’s GLM-5.3. This observation opens the selection of new open models published in the newsletter Latest Open Artifacts.
For a time, custom licenses were very widespread, such as that of Qwen2.5 72B-Instruct or the Llama licenses. DeepSeek initially used a specific license for DeepSeek V3, before R1 switched to the MIT license. In 2025, many Chinese publishers adopted permissive licenses like MIT or Apache 2.0.
Two opposing trends
In 2026, competition among open models is fiercer than ever. Google, with gemma-4-31B-it, and Meta, with Muse-Glimmer-30B, have switched to the Apache 2.0 license. Among leading Chinese players, the trend is reversed. Kimi K3 requires a commercial agreement for providers of inference services (model execution) or fine-tuning (specializing a model on data). MiniMax M3 imposes agreements beyond a revenue threshold and prohibits certain uses.
The latest case is GLM-5.3, which moves from the MIT license, in effect for GLM-5.2 and earlier versions, to a custom license. This license stipulates that if the licensee or any of its affiliated companies operates a Model as a Service-type service and exceeds $10 billion in cumulative revenue over 12 consecutive months, it must undergo a Z.AI security review before any commercial use. According to the text, “the scope and method of this review are reasonably determined by Z.AI”.
The threshold is considered very high compared to other licenses of this type. However, the notion of “affiliates” is not defined in the license, which adds uncertainty. The Chinese version uses the term 关联方, which has a definition in Chinese law. The authors specify they are not legal experts and acknowledge the reasons driving the creation of such licenses.
A selection of new models
Among the cited models, Motif-3, from Motif-Technologies, is published under the MIT license and shows, according to the newsletter, impressive scores for its size, despite limited training resources. dots3-note-prev, from dots-studio, falls under RedNote/Xiaohongshu. According to the source, dots3 achieved a perfect score at IMO 2026 thanks to an internal environment.
GLM-5.3-Flash first appeared as a free “stealth model”, named “Ox-Alpha”, on OpenRouter and OpenCode. Users suggested it could be a model of over 1T from Cursor/xAI, from Gemini, or a new pre-training from open source labs. Hy4-preview, from Tencent, is considered competent but suffers from an “overthinking” problem. According to the newsletter, the trajectory from Hy3-preview to Hy3 suggests the final version could aim for the top ranks of open models.
Finally, Alibaba openly publishes its largest Qwen versions, with Qwen3.8-2.4T-A95B, under a custom license. Its performance is, according to the source, lower than that of other models of its size.
