China and DHS move on open models as DeepSeek ships V4.1 Flash
Original title: Latest open artifacts (#25): DeepSeek V4.1 Flash 1017, American-made models, and new regulatory controls in the US & China
The commerce proposal is not law, and DeepSeek also published DeepInfer, the inference stack it uses to serve tokens.
China’s commerce ministry proposed, but has not enacted, pre-approval before frontier models and some other technologies can be shared abroad, with automatic approval described for open-source technologies. DHS ordered that DeepSeek and its models be barred as vendors in federal civilian procurement; the newsletter says the effect on finetunes and models trained on DeepSeek outputs is unclear. DeepSeek also published DeepInfer, its internal inference stack. The roundup highlights new open weights including DeepSeek-V4.1-Flash-1017, Xiaomi’s MiMo-V2.6-Pro, and Reflection’s Beam-1.0.
Unconfirmed: The MOFCOM text is only a proposal, its intent is disputed, Kimi K3.1 is rumored, and a 90% gross-margin figure is an estimate.
- MOFCOM’s proposal is not law and would auto-approve open-source technologies.
- DHS bars DeepSeek and its models from federal civilian procurement.
- DeepSeek-V4.1-Flash-1017 is a 552B-A16B MIT-licensed model with image inputs.
- Zhipu AI, maker of GLM, is already on the Commerce Entity List.
- Reflection’s Beam-1.0-501B-A23B is released under Apache 2.0.
- Products
- DeepInfer
- Organizations
- Ministry of CommerceDepartment of Homeland Security
- Countries
- ChinaUnited States
From the original article
4 min read at interconnects.ai
The geopolitics of open-weight AI are accelerating again this fall, as we would expect with the risks of all AI systems (open and closed) rising rapidly. We’ve seen many news items in the last week showing both the U. S.
By Florian Brand · interconnects.ai
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.