DeepSeek opened up a cheap model and its inference stack. US federal agencies are now barred from buying it

On October 20, 2026, Florian Brand and Nathan Lambert of Interconnects published issue 25 of Latest open artifacts. The headline is DeepSeek: it released V4.1 Flash 1017 and published its internal inference stack. In the same week, US and Chinese authorities signaled that they want open models under closer watch.

DeepSeek opened both the model and the engine that serves it

According to the authors, DeepSeek has shipped two new iterations since the last roundup. They come in a new size, 552B parameters with 16B active, and also accept image inputs. The model has drawn attention for strong performance at a low price and carries the MIT license. On DeepSWE, an agentic software engineering benchmark, the authors' chart shows DeepSeek dominating the price-performance frontier.

The bigger news is DeepInfer, the internal inference stack DeepSeek published on GitHub. The authors expect it to be far more efficient than vLLM or SGLang for DeepSeek's models. Estimates they link to put DeepSeek's gross margin on token sales at up to 90%. Interconnects says this is the first time a leading lab has shared such tooling.

The roundup also covers other releases. Xiaomi shipped MiMo V2.6 Pro RL with a technical report and some of its training environments. Reflection released Beam 1.0 under Apache 2.0, which the authors call the strongest permissively licensed open American model. Nvidia updated Nemotron 3.5 Super, and Meta released an open version of Muse. The rest of the post is for paid subscribers, so this piece relies on the free preview.

Provenance now matters as much as price

The US Department of Homeland Security (DHS) issued an order barring DeepSeek, both as a company and as a model platform, from being a vendor in federal civilian procurement. The authors note it is unclear how this applies to fine-tuned models or models trained on DeepSeek outputs. Their example is Nvidia's Nemotron, which they say has long relied heavily on DeepSeek models in its training recipes.

China's Ministry of Commerce (MOFCOM), meanwhile, proposed that frontier models and some other technologies need government pre-approval before being shared internationally. It is a proposal, not a law. The authors find the reaction overblown, since the document sets up automatic approvals for open source. Others they trust read it as an attempt to prevent deals like the sale of Manus to Western companies.

For European companies there is a practical takeaway. The US order does not bind them directly. But anyone selling to US agencies or their contractors now has to track where each model comes from and what it was trained on. The MIT license settles copyright questions; it does not settle provenance.

Both moves are narrower than the headlines suggest

DHS only restricts federal civilian purchasing. Private US companies can keep using DeepSeek. China's proposal may not pass in its current form, and it promises automatic approval for open source. The real impact remains uncertain.

The numbers deserve caution too. The 90% margin is an estimate from a post on X, not a figure from DeepSeek. The DeepSWE lead comes from the authors' chart, and one benchmark does not tell you how a model behaves in your own codebase.

The DHS interpretation and inference prices will show who wins

The first signal is whether DHS clarifies that the ban covers models derived from DeepSeek. If it does, US models like Nemotron will have to explain their lineage too.

The second signal is pricing. The authors expect inference providers to adopt DeepInfer quickly, pushing DeepSeek model prices down. The third is the final text of China's rule and, according to the authors, the imminent Kimi K3.1, which for now is only rumored.

Lilith's verdict

DeepSeek published the recipe and the whole kitchen, and Washington told federal agencies not to eat there. Nvidia, however, has already cooked with its ingredients, and nobody has said whether the ban covers its plates too.