1. News
  2. AI models & agents
  3. AI · Open models
AI · Open models2 min read

US and China move to control open AI models as new ones launch

The US barred DeepSeek from federal civilian buying, and China proposed pre-approval for sharing top AI abroad. Tools built on these models could be affected.

By the Dashline deskSource: Interconnects
US and China move to control open AI models as new ones launch
Photo: Interconnects

Both the US and Chinese governments made moves this past week to put some controls around open AI models, the kind whose files anyone can download and run. The US Homeland Security Department barred DeepSeek from federal civilian purchasing. China's Commerce Ministry proposed government pre-approval before top AI is shared abroad. New models kept launching anyway.

What changed

The Homeland Security order bars DeepSeek, as a company, and its models from being a vendor in federal civilian procurement. That means civilian government agencies cannot buy from it. Another Chinese model maker, Zhipu AI, which makes the GLM models, is already on a Commerce Department list that limits its access to American technology.

China's Commerce Ministry published a proposal, not yet a law. It would require frontier AI models and some other technologies to get government approval before being shared internationally. The document says open-source technology would be approved automatically, so the author of the source newsletter found the reaction overblown. Some analysts see it as a framework that could later be used to regulate China's AI industry.

Why prices may move

DeepSeek also released its own internal software for running its models, called DeepInfer. It is likely far more efficient than existing open tools. Estimates say DeepSeek earns up to 90% gross margin selling its AI output on it. Companies that host models are expected to adopt it, which should push prices for DeepSeek models down across the market, or push their margins up.

New models in the mix

  • DeepSeek V4.1 Flash 1017 adds a new size, accepts images, uses the permissive MIT license and is praised for low prices.
  • Reflection's Beam is called the strongest open American model under a permissive Apache 2.0 license.
  • Xiaomi's MiMo V2.6 Pro is the top open model on Artificial Analysis's ranking today.
  • Nvidia's Nemotron 3.5 Super can run at home on a powerful laptop or an RTX Spark.

What is not known yet

It is unclear how the Homeland Security order applies to models fine-tuned from DeepSeek or trained on its outputs. Nvidia's Nemotron models have long leaned on DeepSeek in their training recipes, and it is not clear whether they would be covered. China's proposal could still change before it becomes a rule. Kimi K3.1 is rumored to launch soon.

What to check

If an AI tool you use is built on a Chinese model, ask the vendor which model sits underneath. Federal rules and pricing may shift, especially for anyone selling to the government.

The takeaway

If your AI tools run on Chinese models, ask your vendor which model is underneath, since federal rules and prices may change.

More articles

You should really read these.

Next step

Let’s talk