Latest open artifacts (#25): DeepSeek V4.1 Flash 1017, American-made models, and new regulatory controls in the US & China
Interconnects Florian Brand ● Covered by 4 sources
China and the US both moved to tighten rules around open AI models. DeepSeek also opened up a key internal tool, which could reshape how cheap its models get to run.
Based on reporting by Interconnects, Florian Brand — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Open-weight AI just got dragged deeper into geopolitics. In the past week, both Washington and Beijing moved toward more control over how frontier models are shared, and neither side seems interested in pretending this is a niche issue anymore.
On the Chinese side, the Ministry of Commerce floated a policy proposal that would require pre-approval before frontier AI models and some other technologies could be shared internationally. It is not a law yet. The reaction has been loud, but the document also says open-source technologies would get automatic approvals, which makes the whole thing less black-and-white than the alarmist takes suggest. Even so, the proposal could become the template for future regulation, and some observers think its real target may be broader tech deals, not just AI models.
The US answer was more direct. The Department of Homeland Security issued an order on DeepSeek that blocks the company and its models from being vendors in federal civilian procurement. It is still unclear how that would apply to fine-tuned models or models trained on DeepSeek outputs. That matters, because other systems already appear to lean on DeepSeek in their training recipes. The message, though, is plain enough: the pressure for a domestic open-model stack is growing.
DeepSeek then made the week even stranger by publishing DeepInfer, its internal inference stack. That is not a model or a dataset, but it is the kind of infrastructure top labs usually keep to themselves. The company’s stack is likely more efficient than common open-source tools used for its models, and estimates point to gross margins as high as 90% when it sells tokens on that setup. If inference companies adopt it quickly, prices across the ecosystem could fall, or DeepSeek’s margins could rise. Either way, the release is unusually consequential.
There is plenty else in the latest batch of artifacts: Xiaomi’s MiMo-V2.6-Pro-RL, which leans harder on reinforcement learning; a new DeepSeek-V4.1-Flash-1017 model with image input support and an MIT license; Cloudflare’s clef; NVIDIA’s updated Nemotron; and Reflection’s Beam-1.0-501B-A23B, described as the strongest open American model under Apache 2.0. The common thread is not just better models. It is that the rules around them are hardening at the same time.
My take — AI-written commentary, not fact-checked reporting
The comforting fantasy that open models float above politics is dead. Governments now see them as leverage, industry sees them as infrastructure, and the rest of the market gets to live with the paperwork. The irony is delicious: the more capable open models get, the less open the world around them becomes.
Read more about this at: Interconnects