DHS orders DeepSeek and DeepSeek-based products removed from US federal civilian systems

• Model: DeepSeek-V4.1-Flash-1017. It is a 552B-parameter mixture-of-experts (MoE) model with 1M-token context, released under the MIT license (XenoSpectrum). Checkpoints are on Hugging Face (arXiv paper).
• API price per million tokens, peak (off-peak in brackets):
• Input on a cache miss: $0.30 ($0.15)
• Input on a cache hit: $0.006 ($0.003)
• Output: $1.20 ($0.60)
• What the order does: Federal civilian agencies must remove DeepSeek and affiliate products, plus third-party products built on DeepSeek, and stop buying them immediately (IBTimes UK).

• Architecture, per DeepSeek's V4.1-Flash paper:
• It activates about 16B parameters per token when generating, and about 8B when processing the prompt.
• It is a causal encoder-decoder design aimed at agentic workloads.
• It uses "Compressed Sparse Attention 2," which reuses the KV cache across layers, and stores that cache in FP4. DeepSeek says this cuts the cache to about 890 bytes per token.
• It was pretrained on 45T tokens.
• Evaluation: All scores come from DeepSeek, using harnesses and settings DeepSeek chose. The DeepSWE result is reported at the lowest compute setting (OfficeChai).
• Implementation impact: Open weights mean many teams can self-host or wrap the model, and DeepSeek may be one selectable backend among several in a product. That makes it hard to know which tools fall under the order's "third-party" clause.

• Benchmarks (DeepSeek-reported, not independently checked) (XenoSpectrum):

| Benchmark | Flash-1017 | Opus 5 |

|---|---|---|

| DeepSWE | 77.0 | 74.0 |

| Automation-Bench | 60.5 | 50.3 |

| Agents' Last Exam | 35.4 | 28.6 |

| Apex Bench | 72.1 | 65.8 |

| Terminal-Bench 4.0 | 42.3 | 51.8 |

| Humanity's Last Exam (no tools) | 42.6 | 56.3 |

• Earlier scores differ: Coverage of the original V4.1-Flash reported DeepSWE at 74.2 and Automation-Bench at 54.8 (OfficeChai). When reading any comparison, check which version it refers to.
• DHS's risk case:
• DeepSeek's servers are in China, and Chinese law gives the state broad access to private infrastructure.
• Signup collects personal information and may fingerprint users.
• The models may contain malware or vulnerabilities (video 07:20; IBTimes UK).
• The evidence behind it is public material: DeepSeek's privacy policy, the NowSecure audit and an April 2025 House Select Committee report. No confirmed leak of government data is described (CellCog).

• Earlier device bans: Congress, NASA, the Navy and the Pentagon had banned DeepSeek on their devices by February 2025, and Texas did the same (The Hacker News). The video adds New York, Virginia, Italy, Taiwan, South Korea and Australia (06:02).
• Second use of FASCSA: This is the second exclusion order under the 2018 law. The first, in 2025, targeted Swiss cybersecurity firm Acronis (The Hacker News).
• What's new: The order covers every civilian agency at once, reaches into third-party products, and bans purchasing, not just app installs (video 06:25).

• "DeepSeek implementations" is undefined. One outlet says it is unclear whether DeepSeek models hosted by US providers are covered (The Hacker News). Another says they are (CellCog).
• Exclusions are described differently. The video says the Department of Defense is excluded (02:49). Other coverage says the exclusion is national security systems and secure compartmented information (SCI) systems (IBTimes UK).
• Contractors: The Hacker News says the order applies to contractors. Other coverage says it reaches only agencies.
• No reply yet: DeepSeek has not commented.
• The video's bias: It promotes a paid community and its "agent OS" product (05:18). Its auto-captions also contain errors.

• Vendors selling to the federal government should list which models their products use, including optional and self-hosted backends. They should also be ready to show that DeepSeek is absent or can be removed. Questions go to CISA's FASCSA Order Implementation Team (Tech-Insider).
• Private buyers face no legal ban. The video's main argument is that procurement questions will spread and that being able to swap models is the hedge (09:08; 11:00). That is the commentator's prediction, not a documented trend.
• Cost case: Flash's low price per task makes it attractive on cost alone. Run your own evaluations before using it to replace a frontier model.