PolicyLaw, regulation & governance
DHS Orders Federal Agencies to Remove DeepSeek as the Company Releases V4.1-Flash-1017 and DeepInfer, and China Drafts AI Export Controls
The US Department of Homeland Security ordered federal civilian agencies to remove DeepSeek and products built on it, while DeepSeek released an MIT-licensed V4.1 Flash model and its production inference stack, and China drafted export controls covering frontier model weights.
From the Interconnects article of October 20, 2026 and cited sources: The DHS order was issued under the 2018 federal supply-chain security law. [2] [7] [8]
SOURCES[2] Xenospectrum: DHS Orders Federal Civilian Agencies to… xenospectrum.com[7] Federal Register public inspection: 2026-21447 public-inspection.federalregister.gov[8] PYMNTS: DHS Bans DeepSeek pymnts.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- Original source
- Interconnects article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 20, 2026 article published
- Covered
- Oct 21 edition
- Evidence
- 10 sources across 10 sites
- 1Interconnects: Latest open artifacts (#25)interconnects.ai
- 2Xenospectrum: DHS Orders Federal Civilian Agencies to…xenospectrum.com
- 3SiliconANGLE: DeepSeek open-sources DeepInfersiliconangle.com
- 4Geopolitechs: China's Proposed AI Export Controlsgeopolitechs.org
- 5The Decoder: DeepSeek V4.1 Flash 1017the-decoder.com
- 6Hugging Face: DeepSeek-V4.1-Flash-1017huggingface.co
StrongHow it’s measured
Substack · 20 likes · 5 comments
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
The DHS order was issued under the 2018 federal supply-chain security law. It was signed July 24 and made public October 19–20. It requires federal civilian agencies to remove DeepSeek and also third-party products developed using DeepSeek implementations, so a product's vendor name alone does not decide whether it is covered. [2] [7] [8]
DeepSeek-V4.1-Flash-1017 ships MIT-licensed weights. It takes text and image input and has a 1M-token context window. DeepSeek reports gains over the previous Flash model on DeepSWE, HLE and Automation-Bench. Artificial Analysis, testing independently, ranks it second of 121 open-weight models at about $0.26 per task but finds it notably verbose. [6] [5] [9]
DeepSeek open-sourced DeepInfer, the inference framework behind its own production services, under Apache 2.0. It runs on Nvidia Blackwell and Huawei Ascend 950 chips. It adds speculative decoding and a prefill shortcut but leaves out key-value cache management and load balancing, and an efficient deployment needs roughly 128 GPUs. [3]
China's commerce and science ministries proposed adding 40 technologies to the export-control catalogue. The thresholds include 10^26 training FLOPs, 2 trillion parameters, a global top-ten ranking, or agents that run 20 or more task steps. Open-source releases would be approved automatically but still need registration before export. Public comments close November 15. [4] [10] [1]
Governments on both sides are tightening control while open weights get more capable and cheaper to run.
Read the full assessment
As a result, federal contractors and model buyers need to check where a model came from, not just who sells it, and Chinese labs face a new filing step before in-scope releases.
Sources are gathered and checked. Writing the dossier now.
Executive brief
A US Department of Homeland Security (DHS) order reaches past DeepSeek itself. Federal civilian agencies must also remove "third-party products and services developed using DeepSeek implementations," and the order leaves open whether modified or derivative models are covered. In the same week, DeepSeek released DeepSeek-V4.1-Flash-1017 as an MIT-licensed open-weight model and open-sourced DeepInfer, the inference software behind its own API. China also proposed export controls that cover frontier model weights. Nathan Lambert and Florian Brand's Interconnects roundup describes an open-model ecosystem that is maturing technically while governments on both sides move toward controlling it.
What changed and event timeline
DHS signs DeepSeek order
Order 2026-01 was issued under the Federal Acquisition Supply Chain Security Act of 2018. The 60-day window for agencies to request exceptions counts from this date, not from public release.
China drafts export curbs
China's Ministry of Commerce (MOFCOM) and Ministry of Science and Technology proposed adding 40 technologies to the export-control catalogue. An interim pre-export registration requirement took effect the same day. Public comments close November 15.
DeepSeek-V4.1-Flash-1017 ships
DeepSeek released the model on its website, app and API, and published the weights. It claims stronger results with shorter outputs than the previous version.
More detail
DHS order made public
The notice went to the Federal Register's public inspection site on October 19, with publication scheduled for October 20.
More detail
DeepInfer open-sourced
DeepSeek released the inference framework behind its production services under Apache 2.0. It runs on Nvidia Blackwell GPUs and Huawei Ascend 950 chips.
Capabilities and access
- Model: DeepSeek-V4.1-Flash-1017, the October 17 snapshot. Weights are MIT-licensed (model card).
- Size: sources disagree. Interconnects reports 552B total parameters with 16B active per token, while the Hugging Face card lists 763B.
- Pricing: Artificial Analysis lists $0.30 per million input tokens and $1.20 per million output tokens.
Read the full section
- Model: DeepSeek-V4.1-Flash-1017, the October 17 snapshot. It accepts text and image inputs, has a 1M-token context window and can produce at least 256K output tokens. Weights are MIT-licensed (model card).
- Size: sources disagree. Interconnects reports 552B total parameters with 16B active per token, while the Hugging Face card lists 763B.
- Pricing: Artificial Analysis lists $0.30 per million input tokens and $1.20 per million output tokens.
- DeepInfer: Apache 2.0, with a companion library, deepseek-recipe, that converts API requests into the model's token format (SiliconANGLE).
Technical analysis for researchers and developers
- DeepInfer features: speculative decoding through a component called DSpark, which drafts five tokens at once and uses a confidence head to pick likely ones (DeepSeek reports 4–7% higher throughput).
- Deployment limits: key-value cache management and load balancing are not included.
- Lineage: DeepInfer started as an early fork of vLLM.
Read the full section
- DeepInfer features: speculative decoding through a component called DSpark, which drafts five tokens at once and uses a confidence head to pick likely ones (DeepSeek reports 4–7% higher throughput). A prefill shortcut that sends only the final 128 prompt tokens through the decoder layers roughly halves prefill compute (SiliconANGLE).
- Deployment limits: key-value cache management and load balancing are not included. An efficient deployment needs about 64 GPUs for prefill plus 64 for decoding, so this is cluster-scale software. Hopper and older Nvidia GPUs are not supported, and deterministic output works only on Nvidia hardware.
- Lineage: DeepInfer started as an early fork of vLLM.
- Prompt formatting: the model card ships no Jinja chat template. Developers must use DeepSeek's Python encoding code or deepseek-recipe (model card).
Claims and evidence
- Benchmarks (DeepSeek's own numbers): compared with the previous V4.1-Flash, DeepSWE v1.1 rose from 74.2 to 77.0, HLE from 36.8 to 42.6 and Automation-Bench from 54.8 to 60.5 (model card).
- Independent evaluation: Artificial Analysis scores the model 45 on its Intelligence Index, second among 121 open-weight models, at $0.26 per task.
- DeepInfer economics: one cluster serves 155 trillion tokens a day, and off-peak revenue reaches $38 per GPU-hour.
Read the full section
- Benchmarks (DeepSeek's own numbers): compared with the previous V4.1-Flash, DeepSWE v1.1 rose from 74.2 to 77.0, HLE from 36.8 to 42.6 and Automation-Bench from 54.8 to 60.5 (model card). DeepSeek also claims about $0.20 per DeepSWE task, against $2 to over $10 for Claude Opus 5 and GPT-6 Astra. The Decoder notes that independent benchmarks hadn't come out yet.
- Independent evaluation: Artificial Analysis scores the model 45 on its Intelligence Index, second among 121 open-weight models, at $0.26 per task. It also ranks the model as verbose: 210M output tokens against a 140M median. That cuts against DeepSeek's "shorter outputs" claim.
- DeepInfer economics: one cluster serves 155 trillion tokens a day, and off-peak revenue reaches $38 per GPU-hour. All published benchmarks come from DeepSeek's own tests on Nvidia hardware (SiliconANGLE). Interconnects' estimate of up to 90% gross margin has no independent corroboration.
Context and prior work
- DeepInfer bundles kernels and communication libraries DeepSeek had already open-sourced: FlashMLA, DeepGEMM and DeepEP (SiliconANGLE).
- Zhipu AI, which makes the GLM models, is already on the US Commerce Department's Entity List (Interconnects).
- The same Interconnects issue highlights other releases. NVIDIA released Nemotron-3.5-Super-VL along with most of its training data.
Read the full section
- DeepInfer bundles kernels and communication libraries DeepSeek had already open-sourced: FlashMLA, DeepGEMM and DeepEP (SiliconANGLE).
- Zhipu AI, which makes the GLM models, is already on the US Commerce Department's Entity List (Interconnects).
- The same Interconnects issue highlights other releases. Xiaomi's MiMo-V2.6-Pro came with a tech report and some of its reinforcement-learning training environments. Reflection's Beam-1.0-501B-A23B (Apache 2.0) is, in the authors' view, the strongest permissively licensed American open model. NVIDIA released Nemotron-3.5-Super-VL along with most of its training data.
Limitations, safety and contested findings
- DHS rationale: DeepSeek's models "present multiple threat vectors," including possible data access by Chinese authorities and cybersecurity weaknesses (PYMNTS).
- How serious is China's proposal: Interconnects calls the reaction overblown, because open-source releases get automatic approval.
- China's thresholds: foundation models trained with 10^26 FLOPs, having 2 trillion parameters, or ranked in the global top ten.
Read the full section
- DHS rationale: DeepSeek's models "present multiple threat vectors," including possible data access by Chinese authorities and cybersecurity weaknesses (PYMNTS). Interconnects asks whether models trained on DeepSeek outputs, such as Nvidia's Nemotron, could be caught by the order.
- How serious is China's proposal: Interconnects calls the reaction overblown, because open-source releases get automatic approval. Analysts note it still adds a filing step before any open release that falls in scope (dev.to). Geopolitechs reads it as MOFCOM gathering data before deciding final controls.
- China's thresholds: foundation models trained with 10^26 FLOPs, having 2 trillion parameters, or ranked in the global top ten; agents that run 20 or more task steps (Geopolitechs).
Business and practitioner implications
- Federal contractors: check whether any product was "developed using DeepSeek implementations."
- Inference providers: DeepInfer could cut serving costs for DeepSeek models, but only for operators with Blackwell or Ascend clusters at the 128-GPU scale.
- Model buyers: Beam (Apache 2.0) and Nemotron are US-origin open alternatives that avoid this procurement risk.
Read the full section
- Federal contractors: check whether any product was "developed using DeepSeek implementations." Under the order, the vendor's name alone does not settle whether a product is covered (Xenospectrum).
- Inference providers: DeepInfer could cut serving costs for DeepSeek models, but only for operators with Blackwell or Ascend clusters at the 128-GPU scale.
- Model buyers: Beam (Apache 2.0) and Nemotron are US-origin open alternatives that avoid this procurement risk.
- Chinese labs: even open releases that fall under the draft thresholds now require MOFCOM registration first.