DeepSeek releases MIT-licensed V4.1-Flash-1017 as DHS issues federal civilian exclusion order
DeepSeek released its MIT-licensed V4.1-Flash-1017 coding model with an 890-byte per-token KV cache on October 17, as DHS published an October 20 FASCSA order barring DeepSeek across federal civilian IT.
Two developments over four days sharpened the split between open-weight technical capability and US procurement law.
On October 17, DeepSeek released V4.1-Flash and the agent-tuned V4.1-Flash-1017 under the MIT License. The 763-billion-parameter Mixture-of-Experts architecture activates 8 billion parameters per token during prefill and 16 billion during decoding across a 1-million-token context window, using Cross-Stage Attention (CSA2) to compress its FP4 KV cache to 890 bytes per token. On vendor evaluations, V4.1-Flash-1017 raised Terminal-Bench 4.0 pass@1 from 31.2% to 42.3% and scored 77.0% on DeepSWE v1.1. [1][2]
On October 20, the Department of Homeland Security published formal notice of Exclusion and Removal Order #2026-01 (91 FR 66346, FR Doc. 2026-21447, Docket DHS-2026-0628) issued under the Federal Acquisition Supply Chain Security Act (FASCSA). The order, signed July 24, 2026, requires all federal civilian executive branch agencies to remove DeepSeek products and bars procurement not only of DeepSeek software but also of downstream third-party products and services built upon covered DeepSeek implementations. [3][4]
For engineering leaders, the technical and legal boundaries must be evaluated together. All 763 billion parameters still require multi-node GPU memory even when per-token KV cache drops to 890 bytes, and any software vendor selling into US federal civilian agencies must ensure DeepSeek weights, hosted endpoints and fine-tunes remain excluded from its federal delivery stack. [1][3][4]
Take it further
Read the DHS FASCSA exclusion order on Federal Register For federal contractors and compliance teams. Details covered DeepSeek entities, applicable definitions and downstream-implementation restrictions across federal civilian IT.
Inspect DeepSeek V4.1-Flash-1017 architecture and MIT weights For non-federal commercial and research teams. Check memory requirements for the full 763B MoE weights before self-hosting.
Sources & details 4 sources
Discuss in your voice app
Take the reporting, visual explanations and original sources into a conversation. Copy the brief, paste it into your preferred assistant, then start its voice mode.
Paste the brief into your voice app to discuss this reporting.Reporting sources
- DeepSeek · V4.1-Flash-1017 model card and MIT licence
Primary October 17 repository: 763B total parameters, 8B prefill / 16B decode active, 890-byte FP4 KV cache, 42.3% Terminal-Bench 4.0 and 77.0% DeepSWE v1.1 (vendor-reported).
- DeepSeek · V4.1-Flash base repository
Base October 17 model card reporting 31.2% Terminal-Bench 4.0 pass@1 and 75.0% DeepSWE v1.1.
- Federal Register · DHS FASCSA exclusion and removal order
FR Doc. 2026-21447 (91 FR 66346), published October 20, 2026, Docket DHS-2026-0628. Formal notice of Exclusion and Removal Order #2026-01 under 41 U.S.C. 1323.
- Security Magazine · Federal civilian ban on DeepSeek
Independent October 20 reporting corroborating the DHS FASCSA removal order and its coverage of downstream third-party services.
Images & figures
V4.1-Flash-1017 raises Terminal-Bench 4.0 pass@1 to 42.3%, while DHS bars DeepSeek across federal civilian IT.
Euangelion · original data figure · Source · CC BY 4.0 · Download figure