The open model license landscape is splitting along geographic lines. Google and Meta both moved to Apache 2.0 in 2026. Chinese frontier labs are moving the opposite direction. Kimi K3 requires commercial agreements for inference and fine-tuning services. MiniMax M3 adds revenue thresholds and prohibited use cases. Zhipu's GLM-5.3 dropped MIT, which covered GLM-5.2 and all prior versions, for a custom license that requires passing Z.AI's security review before any commercial use once a licensee's aggregate revenue exceeds $10 billion USD over any consecutive 12-month period.
The $10 billion threshold sounds permissive until you read the fine print. The license never defines 'affiliates,' which injects legal uncertainty for any large organization considering deployment. The license is published in both English and Chinese, and the Chinese text uses a term, 'guanlianfang,' that carries a specific definition under Chinese accounting law. That dual-language structure, with one version anchored to a foreign legal system, is the detail most adopters will miss and the one most worth reading carefully in the original.
Two models stand out in this edition beyond the license discussion. Motif-3 from Motif-Technologies ships under MIT and posts strong benchmark scores relative to its parameter count, continuing a trajectory that started with the 2.6B model in 2025. dots3-note-prev from dots-studio, the model lab connected to RedNote/Xiaohongshu, achieved a perfect score at IMO 2026 using an internal harness. Both are worth tracking. The full piece covers benchmark breakdowns and the complete license clause text.
[READ ORIGINAL →]