DeepSeek releases V4.1 Flash 1017 model and open inference stack amid US/China regulatory shifts

DeepSeek released DeepSeek-V4.1-Flash-1017, a model with 552B total parameters and 16B active parameters that supports image inputs. The model is distributed under the MIT license, continuing a trend toward permissive licensing in open-weight releases.

The company also open-sourced its internal inference stack, DeepInfer, which is positioned as more efficient than existing open-source tools like vLLM and SGLang. Estimates cited in the report suggest DeepSeek achieves up to 90% gross margins on token sales using this stack, implying potential cost reductions or margin increases for users and the company alike.

China’s Ministry of Commerce (MOFCOM) proposed a policy requiring pre-approval for sharing frontier AI models internationally, though the document indicates open-source technologies may qualify for automatic approval. The proposal is framed as a regulatory framework that could expand in scope, though some analysts interpret it as more directly aimed at preventing the sale of Chinese startups to Western firms.

The US Department of Homeland Security (DHS) issued an order restricting DeepSeek, as a company, and its models from federal civilian procurement. The order does not clarify whether fine-tuned or derivative models (such as Nvidia’s Nemotron series) are included in the restriction.

Separately, the report notes that other Chinese model developers, including Zhipu AI (makers of GLM), are already listed on the US Entity List, which limits their access to American technology. The report also highlights the release of Beam-1.0-501B-A23B by Reflection, described as the strongest open American model released under a permissive license (Apache 2.0), and positions it as token-efficient.