About this Resource
<p><i><span style="font-size: 16px;">Explore how new government regulations and AI licensing restrictions could reshape open-source AI, affect future model releases, and change how developers and everyday users access powerful AI tools.</span></i></p><h2><br></h2><h2>Open Source AI Is Quietly Closing: New Rules and Licenses Could Change Everything</h2><p><span style="font-size: 16px;">Open-source AI has made powerful technology more accessible, allowing people to download AI models, run them on their own computers, and build products without relying entirely on expensive services. However, new Chinese export rules and increasingly restrictive model licenses could make access to the next generation of AI more complicated. Two major changes are emerging, China is considering new controls on exporting advanced AI models, while several AI companies are introducing licensing conditions that could limit how businesses use and distribute their models.</span></p><h2><br></h2><h2>China’s New AI Export Rules</h2><p><span style="font-size: 16px;">On October 16, 2026, China’s commerce and science ministries published a draft proposal to update the list of technologies requiring export permission. The proposal includes AI technologies and the core components of large AI models, including their weights. <b>Model weights</b> are the files containing the learned information that allows an AI model to generate responses. When developers release these files publicly, people can download and run the models themselves rather than accessing them only through a website or API.</span></p><p><span style="font-size: 16px;"><br></span></p><p><span style="font-size: 16px;">Under the proposed rules, an AI model could fall under export controls if it meets any one of three conditions:</span></p><p></p><ul><li><span style="font-size: 16px;"><b>Model size:</b> It has at least two trillion parameters.
</span></li><li><span style="font-size: 16px;"><b>Training compute:</b> It requires (10^{26}) or more computational operations to train.</span></li><li><span style="font-size: 16px;"><b>Global ranking:</b> It ranks among the world’s top 10 models in a qualifying category, such as coding, safety, performance, or usage.</span></li></ul><p></p><p><span style="font-size: 16px;"><br></span></p><p><span style="font-size: 16px;">The third condition is particularly important because it could affect models that are smaller than two trillion parameters but highly successful in a specific area. The draft proposes that restricted models would require export licensing, while new open-source releases could receive automatic licensing. AI companies could still offer access through websites, apps, and APIs, meaning the main concern is whether users can freely download the model files. The draft’s public-comment period ends on November 15, 2026, and a date for final rules has not yet been announced.</span></p><h2><br></h2><h2>Could Alibaba’s Qwen Models Be Affected?</h2><p><span style="font-size: 16px;">Alibaba’s upcoming Qwen models are an important example of why these rules matter. Qwen 4 being in training, with Alibaba’s longer-term roadmap reportedly targeting models containing between five and ten trillion parameters. Those future models would exceed the proposed two-trillion-parameter threshold. However, the rules would not automatically mean that every future Qwen release would be blocked. The proposal includes an automatic licensing route for open-source releases, and its final requirements remain uncertain. There is also an important distinction between existing models and future releases. Models already publicly available through open-source releases or research papers are described in the draft as exempt from the proposed controls. This means that people who have already downloaded a model could continue using their local copy, even if access to a future version became more restricted.</span></p><h2><br></h2><h2>Why the Top-10 Ranking Rule Matters</h2><p><span style="font-size: 16px;">The proposed ranking condition could potentially affect more models than the size threshold alone. A model might qualify because it performs exceptionally well at coding or another specialised task, even if it is nowhere near two trillion parameters. The draft also considers usage, which creates further uncertainty. Chinese AI models accounted for a substantial share of requests on OpenRouter during the week of October 12, 2026, as highlighted in the video. However, important questions remain unanswered. The draft does not clearly establish which rankings will count, how frequently a model’s position will be assessed, or what happens if its ranking changes. Until these details are clarified, it is difficult to know exactly which future models could be affected.</span></p><h2><br></h2><h2>AI Companies Are Adding Restrictions to Their Own Licenses</h2><p><span style="font-size: 16px;">Government regulation is only one part of the problem. Some AI companies are also changing the licenses attached to their downloadable models. Traditional licenses such as MIT and Apache generally allow broad use, including commercial applications, subject to their terms. These licenses have helped businesses build AI products and offer affordable model-hosting services. Newer licenses, however, can introduce additional requirements.</span></p><p><span style="font-size: 16px;"><br></span></p><p><span style="font-size: 16px;">The video highlights several examples:</span></p><p></p><ul><li><span style="font-size: 16px;">Kimi K3: Model-hosting businesses earning more than $20 million in annual revenue must negotiate a separate agreement with Moonshot AI.
</span></li><li><span style="font-size: 16px;">Muse Spark Open: Its license includes a similar $20 million threshold for model-as-a-service businesses and requires products to acknowledge that they were built with Muse Spark Open.
</span></li><li><span style="font-size: 16px;">Qwen: Certain versions have additional licensing requirements, including separate agreements for some larger commercial hosting operations.
</span></li><li><span style="font-size: 16px;">GLM: Newer versions introduce additional requirements for certain hosting providers, including a security review for businesses above a specified revenue threshold.
</span></li><li><span style="font-size: 16px;">MiniMax: Newer licensing terms introduce branding and permission requirements for some commercial uses.</span></li></ul><p></p><p><br></p><p><span style="font-size: 16px;">These conditions do not necessarily prevent ordinary users from downloading and running the models. Instead, they can make it more difficult for larger companies and startups to offer the models as paid services without negotiating additional agreements. For anyone building a commercial AI product, checking the license is therefore becoming just as important as checking the model’s performance.</span></p><h2><br></h2><h2>DeepSeek Remains an Important Open Alternative</h2><p><span style="font-size: 16px;">While some major AI labs are introducing more restrictive licenses, DeepSeek is highlighted as an exception. Its models continue to use the permissive MIT license, which offers broad freedom to use, modify, and distribute software, subject to the license terms. Smaller open models, including Qwen 3.8 27B and other Flash variants, are also presented as important options for people who want to run AI locally. These models can be particularly useful because they are more practical to run on personal computers with suitable hardware. For individuals who value privacy, local control, and independence from cloud subscriptions, downloading and keeping a copy of a permissively licensed model can help preserve access even if future releases become more restricted.</span></p><h2><br></h2><h2>The United States Is Also Taking Action</h2><p><span style="font-size: 16px;">The video also discusses a US Department of Homeland Security order published on October 20, 2026, concerning DeepSeek. The order restricts DeepSeek from federal civilian government agencies, including certain third-party products and services using its implementations. However, the restriction described in the video applies to government agencies. It does not establish a general ban on ordinary people downloading or running DeepSeek models on their own computers. For most individual users, the more immediate issue remains access to future models and the terms attached to the models they choose.</span></p><h2><br></h2><h2>Should You Switch to American Open Models?</h2><p><span style="font-size: 16px;">Not necessarily. Chinese models remain highly competitive, particularly among the strongest open-weight AI systems. Switching simply because of concerns about future restrictions could mean giving up useful capabilities without gaining a clear advantage. However, having a backup is sensible. Depending on your hardware and requirements, alternatives mentioned in the video include Google’s Gemma 4, OpenAI’s GPT-OSS models, and NVIDIA’s Nemotron family. Larger systems may also be suitable for users with more powerful computers. The important step is to <b>test an alternative before you need it</b>. Check its performance on your tasks, hardware requirements, license, and compatibility with your existing tools. A working backup gives you flexibility if a preferred model becomes unavailable or its licensing conditions change.</span></p><h2><br></h2><h2>What Should You Do Now?</h2><p><span style="font-size: 16px;">If you rely on open-source AI, there are several practical steps worth taking:</span></p><p></p><ul><li><span style="font-size: 16px;"><b>Download the models you depend on.</b> Keep local copies of useful model weights rather than assuming every future version will remain freely available.
</span></li><li><span style="font-size: 16px;"><b>Save the license files.</b> Store the relevant license alongside each model so you know which terms applied to your copy.
</span></li><li><span style="font-size: 16px;"><b>Prioritise permissive licenses. </b>MIT and Apache-licensed models can offer greater flexibility than custom licenses, although their exact terms should still be checked.
</span></li><li><span style="font-size: 16px;"><b>Test a backup model.</b> Make sure an alternative works on your hardware and performs well enough for your needs.
</span></li><li><span style="font-size: 16px;"><b>Review commercial restrictions.</b> If you operate a startup or model-hosting service, check whether your intended use requires additional permission or a separate agreement.</span></li></ul><p></p><p><span style="font-size: 16px;"><br></span></p><p><span style="font-size: 16px;">The video specifically highlights permissively licensed options such as Qwen 3.8 27B, GLM Flash variants, and DeepSeek V4.1 Flash as models worth considering, subject to verifying the license attached to the exact version you download.</span></p><h2><br></h2><h2>The Warning Sign to Watch</h2><p><span style="font-size: 16px;">One of the clearest indicators of whether China’s proposed rules are affecting AI development will be the gap between a model’s API launch and the public release of its downloadable weights. The video points to MiniMax M3.1 Flash as an example to watch. Its API became available on October 19, 2026, while its model weights were expected to follow later. A delay alone does not prove that export controls are responsible. However, if future Chinese models repeatedly become available through APIs before their downloadable weights are released, that pattern could indicate a meaningful change in how open models are distributed. Alibaba’s upcoming Qwen releases will also be important to watch.</span></p><h2><br></h2><h2>Final Thoughts</h2><p><span style="font-size: 16px;">Open-source AI is not disappearing overnight, and existing public models are not automatically being taken away from people who have already downloaded them. The bigger concern is what happens next. Government export controls could affect the release of future advanced models, while custom commercial licenses could limit how businesses provide access to models that remain publicly available. For everyday users, the most practical response is to preserve access to useful models, understand their licenses, and maintain a tested alternative. <b>The future of open-source AI may depend not only on which models developers create, but also on whether people are still allowed to download, modify, and build businesses around them.</b></span></p>