The video discusses recent regulatory changes in China affecting the distribution of large AI models, focusing on Alibaba’s upcoming Qwen 4 model. Since October 16th, China has introduced draft export controls requiring permission to download AI models with over 2 trillion parameters, or those considered globally leading in performance or usage. This move places a gate on the release of large open-weight models, meaning labs like Alibaba must seek government approval before allowing downloads of such models. However, Chinese labs can still offer these models via APIs or apps, with the restriction primarily targeting direct downloads of the model files (weights).
The draft rules define three triggers for regulation: model size (2 trillion parameters or more), compute used in training, and global ranking in performance or usage. While only a few Chinese models currently exceed the size threshold, the global ranking and usage criteria could apply to smaller models as well, potentially broadening the scope of regulation. Alibaba’s Qwen 4, expected to have 5 to 10 trillion parameters, will almost certainly fall under these controls. Despite this, existing smaller models like Qwen 3.8 with 27 billion parameters remain unaffected and widely available for personal use.
Alongside government regulations, AI labs themselves have introduced new licensing restrictions on their models. For example, Kimmy K3 and Meta’s Muse Spark Open models include clauses requiring separate agreements for commercial use exceeding $20 million in revenue, and mandates to credit the original model in products. These custom licenses mark a shift away from fully open-source models under permissive licenses like MIT or Apache, aiming to control commercial exploitation and hosting of large AI models. This trend limits the free tier of open models and impacts startups and hosting services that rely on open weights.
In the US, the Department of Homeland Security has restricted the use of DeepSeek models within federal civilian agencies, though this does not affect the general public or private sector. Unlike China, the US currently does not require permission to download or use open-weight models, but legislative proposals could change this in the future. Despite these geopolitical and regulatory challenges, Chinese open-weight models still dominate the top rankings in quality and usage, making them the preferred choice for many users and developers worldwide.
The video concludes with advice for users to save copies of existing open-weight models released under permissive licenses, as future versions may come with stricter controls or licenses. It also suggests monitoring the timing of model weight releases relative to API availability as an indicator of how strictly China’s new gate is being enforced. Overall, while the era of fully open large AI models may be narrowing due to regulatory and licensing changes, smaller models and existing releases remain accessible, and the landscape continues to evolve rapidly.
