DeepSeek V4.1 Flash lands as US and China tighten AI controls

Nathan Lambert and Florian Brand's Interconnects roundup catalogues a fresh 552B-parameter DeepSeek release landing in the same week regulators in Washington and Beijing both moved on how open models cross borders.

DeepSeek-V4.1-Flash-1017 is a sparse mixture-of-experts model at a "new size (552B-A16B) which also support image inputs," shipping under the MIT license. The post calls it a leader on the DeepSWE agentic coding benchmark, and separately flags that DeepSeek has released DeepInfer, the internal inference stack it runs in production. "Estimates point to DeepSeek making up to 90% gross margin when selling tokens on this inference stack," the authors write.

The regulatory moves on either side land harder than the weights themselves. A new Department of Homeland Security order "restricts DeepSeek itself, as a company, and DeepSeek's models, as a technology platform, from being a vendor in federal civilian procurement," though the roundup flags open uncertainty around finetunes and distilled derivatives. On the other side, a MOFCOM proposal would require pre-approval before frontier models can be shared internationally, with one explicit carve-out: "open-source technologies will be set up for automatic approvals."

On the American side, Reflection's Beam-1.0-501B-A23B is called the "strongest open American model released under a permissive license (Apache 2.0)," paired with a long-awaited NVIDIA Nemotron 3.5 refresh and Xiaomi's MiMo-V2.6-Pro-RL, which the authors say "is the top open model on Artificial Analysis's ranking today."