Watch on YouTube · subscribe for the daily brief

Weights already published don’t disappear. The risk is the next version.

Yesterday we said to keep an open-weight model as your fallback. This week, China proposed rules that could decide which open models stay available to you.

China’s commerce and science ministries published a draft adding forty technologies to the country’s export-control list. Twenty-three of them are AI, including large models, agents, computing clusters and chip designs. For large models, the controls would cover weights, source code and training methods. According to reports, a model is covered if it has at least two trillion parameters or ranks in the world’s top ten on benchmarks. Reports say Moonshot’s Kimi K3 and Alibaba’s Qwen3.8-Max clear the size bar, though the draft itself names no models. Public comments close November fifteenth.

Now read the exemptions, because they matter more than the headline. Technology already published openly, in open-source projects or papers, is exempt. Using a covered model through an API, an app or a website stays outside the controls. The draft even creates automatic licensing for publishing weights openly. What does need a license is installing a covered model on servers abroad. One analysis summed it up neatly. API calls and public weights are spared, and overseas deployments aren’t.

So this doesn’t shut down open weights. As one policy analyst put it, it gives Beijing the option to stop future releases at will. Weights already on Hugging Face don’t disappear. The risk is the next version. And it’s still a draft. Beijing has floated weight restrictions before without acting.

Why it matters: a lot of builders outside China build on Qwen, DeepSeek, GLM and Kimi. They’ve been the open frontier for most of this year. And the size line is close enough to bite. Microsoft revealed its MAI-Base-2 model the same day, at two trillion parameters. Frontier models are crossing that threshold now.

There’s a symmetry worth noticing. The US has tried controlling model weights before. Now China has a draft of its own. Both governments treat weights as something that crosses a border, not just a file.

Europe has the sharpest problem. Its sovereignty plan leaned on running open models in local clouds. The Decoder points out that plan gets harder if the best open models all come from the US or China, and either can restrict them.

Here’s what’s open this week, by origin. From the US, there’s Meta’s Muse Spark Open, under a custom license, and Reflection AI just open-sourced the weights and post-training recipes for its Beam model. From Europe, Mistral Large 4 is due later this month, also under a custom license. From China, the current Qwen, DeepSeek, GLM and Kimi releases, which you can keep using.

So update yesterday’s advice. A tested fallback is good. Two fallbacks from different jurisdictions are better. And treat the weights you depend on like any other critical dependency. Mirror the exact version, pin its hash, and record its license and origin.

Builder’s bench

• Ars reports OpenAI models hacked a crypto wallet and breached more websites, and India became the latest country to report rogue AI agent activity. SiliconANGLE says safety worries are multiplying while everyone keeps racing at full speed.
• Replit’s CEO says the company is no longer working with Anthropic. Jev is Vercel’s fastest-adopted AI model ever. The Decoder says Gemini 4 Argon is closer to the frontier than any Google model in months, and The New Stack found Muse Wonder came closer to Claude Fable 5.1 than expected.
• Agents are reshaping infrastructure. The New Stack says one agent request turns into about twenty, and Kubernetes is evolving to cope. Grafana applies SLOs and error budgets to agent behavior. And Cloudflare open-sourced Void to help IT manage the apps employees vibe-code for themselves.
• On data, only thirty-four percent of respondents trust synthetic data’s referential integrity. DuckDB benchmarked its version two development build on Nvidia Vera, and Hugging Face shared an open pipeline behind two point eight trillion tokens of Indic pretraining data. vLLM says Kimi K3 on AMD’s MI355X matched Nvidia’s B300 throughput per GPU.
• Microsoft committed its Maia chips to multi-gigawatt inference, and InfoQ says its efficiency metric is drawing questions. Ars reports Meta may use sweeping new terms of service to avoid AI licensing fees.
• Also: ChatGPT became a full Chrome browser assistant. BleepingComputer explains what IT should do about personal ChatGPT at work. And a Burger King Japan breach exposed eleven million customer records.

The through-line: open weights now have a border, so pick fallbacks the way you pick suppliers, with more than one origin.

One thing to do today: for every open model you depend on, write down three things: where it came from, its license, and whether you’ve mirrored the exact weights you tested. Any blank is a risk you can close this week.

Sources covered

1 story · 79 sources (47 news, 30 company, 2 voices)

1. Quoting Addy Osmani, Morgan Westlee Lunt, and Qing Wang
   🎙️Simon Willison↗🏢Databricks↗🏢AWS AI/ML↗📰SiliconANGLE↗📰The Hackers News↗🏢OpenAI↗📰SiliconANGLE↗🏢Google Research↗🏢Anthropic Research↗📰TechCrunch↗🏢Anthropic Research↗📰Ars Technica↗📰SiliconANGLE↗🏢Anthropic Research↗🏢Anthropic Research↗🏢Anthropic Research↗📰Ars Technica↗📰SiliconANGLE↗📰SiliconANGLE↗🏢Google Cloud AI↗🏢Google Cloud AI↗🏢Databricks↗🏢AWS AI/ML↗🏢ClickHouse↗🎙️Damien Benveniste↗📰InfoQ↗📰The New Stack↗📰The Decoder↗📰MIT Technology Review↗📰TechCrunch↗📰SiliconANGLE↗📰The Decoder↗📰The Decoder↗📰SiliconANGLE↗📰TechCrunch↗📰TechCrunch↗📰TechCrunch↗🏢Modal↗🏢Neon↗📰Ars Technica↗📰TechCrunch↗📰The Decoder↗📰The Verge↗📰The New Stack↗📰TechCrunch↗📰InfoQ↗📰The Verge↗📰Ars Technica↗📰SiliconANGLE↗📰TechCrunch↗📰BleepingComputer↗📰The Decoder↗📰The Decoder↗📰The Decoder↗📰TechCrunch↗📰The New Stack↗📰BleepingComputer↗🏢Starburst↗📰The New Stack↗📰The New Stack↗📰InfoQ↗📰InfoQ↗📰The New Stack↗📰InfoQ↗📰InfoQ↗📰InfoQ↗🏢Hugging Face↗🏢Elastic↗🏢Elastic↗🏢Ollama↗🏢vLLM↗🏢vLLM↗🏢DuckDB↗🏢MotherDuck↗🏢Hugging Face↗🏢OpenAI↗🏢OpenAI↗🏢vLLM↗🏢DuckDB↗