DeepSeek-V4.1-Flash-1017 arrives with improved benchmarks just as the U.S. government bans it from agencies — what enterprises should do

DeepSeek keeps giving businesses reasons to try its low-cost, high-performing, open source AI models. The U.S. government, though, is saying 'no dice.'

On October 17, the popular Chinese AI startup released DeepSeek-V4.1-Flash-1017, promising better coding and faster completion of complex tasks.

Two days later, the U.S. Department of Homeland Security disclosed an order from the summer directing federal civilian agencies including the State Department, Treasury Department and Internal Revenue Service (IRS) to exclude and remove all DeepSeek technology, including implementations that other companies supply.

For a business selling software to the federal government, this raises an immediate, practical question: should you be using DeepSeek in any U.S. federal-facing products, services or solutions?

The answer is an unfortunate “no.”

Civilian agencies now operate under the DHS order, and defense contractors already work under a statute that bars DeepSeek from Pentagon contract work.

However, if a business is offering purely commercial work that never touches a government contract, then DeepSeek remains a viable option — even if the business also serves the government, provided that DeepSeek stays walled off from that government work.

No federal rule bars a U.S. company from using DeepSeek with its private customers. One caveat: sending export-controlled technical data to DeepSeek’s own China-hosted services, whether its app, website or application programming interface (API), can count as an export that requires a license. Downloading the weights, or running the model through a U.S.-based cloud provider or hyperscaler, avoids that particular problem.

That being said, the larger question around U.S. businesses employing any Chinese model remains. Put another way: if the government is doing this to DeepSeek, what’s to stop it from turning around and issuing similar orders or expansions of the ban against rivals such as Alibaba (with its Qwen models), MiniMax (M), Xiaomi (MiMo) and Moonshot (Kimi K)?

Very little in current law, and the signs suggest Washington isn’t finished. Senators have already written a defense-bill provision that would extend the Pentagon’s ban to ten more Chinese AI companies and their derivative models, and the White House spent the summer weighing limits on Kimi and Qwen.

Since the DeepSeek order took effect months before the public heard about it, companies that sell AI to the government shouldn’t wait for the next one to name their model. They should start de-risking now.

What enterprises with government contracts or potential business should do now

Government suppliers face the most immediate decision. The DHS order covers civilian executive agencies, but excludes specified national security systems, such as systems supporting military command or intelligence activities, which fall under separate defense and intelligence authorities. Suppliers cannot treat that jurisdictional distinction as general permission to use DeepSeek for national security work.

Still, here's our proposed checklist if you're contracting with the federal government (note, we're not lawyers and this is not and should not be construed as legal advice):

1. Find every Chinese model in your stack: Inventory your products, employee tools and subcontractors' software, and make every AI vendor name its base models, including fine-tunes and derivatives, and where it runs them.

2. Wall off government work today: Remove DeepSeek from anything you deliver to agencies or use to perform federal contracts, and keep Chinese-developed models out of defense work entirely unless your counsel clears a specific use.

3. Stop adding new dependencies: Don't build new government-bound products on Chinese base models, even where today's rules allow it, because the pending defense provisions would reach derivatives and a product outlives a legislative session.

4. Read your contracts and watch the clock: Check SAM.gov for new supply-chain orders at least every three months, as FAR 52.204-30 requires, talk to your contracting officer before you assume a rule doesn't apply, and know the three-business-day reporting deadline cold.

5. Engineer for portability: Call models by explicit version names, keep at least one alternative tested on your real tasks for accuracy, speed and total cost, and design so that a forced switch takes a configuration change rather than a rebuild, whichever government flips the switch.

6. Keep commercial options open, deliberately: If DeepSeek saves real money on commercial lines, keep it there behind clear technical and contractual walls, and budget for the duplicate testing and oversight that separation demands.

Why DeepSeek-V4.1-Flash-1017 is such an upgrade for enterprises and developers

The business case for considering DeepSeek's new model upgrade remains substantial. American businesses can download its MIT-licensed model weights—the files containing what the model learned—and run or adapt it themselves.

The company’s benchmark table reports a DeepSWE v1.1 software engineering score of 77.0, up from 74.2, versus 74.0 for Claude Opus 5. Terminal-Bench 4.0 rises from 31.2 to 42.3, while Opus 5 remains ahead at 51.8. These are DeepSeek’s reported comparisons, rather than independent confirmation of superiority across enterprise work.

Third-party AI evaluation firm Artificial Analysis scores V4.1-Flash-1017 at 45, up from 39 for its predecessor, while recording a decline in average evaluation cost from $0.49 to $0.26 per task — that is, about half the cost to complete a task as its already cheap predecessor.

The new model costs $0.30 per 1 million uncached input tokens and $1.20 per million output tokens through DeepSeek's API at peak times, with half-price rates off-peak.

DeepSeek says peak hours are 01:00–04:00 and 06:00–10:00 UTC Monday–Friday, excluding Chinese public holidays.

For businesses running it on their own hardware or on U.S. clouds, the costs will obviously vary.

Notably, DeepSeek-V4.1-Flash-1017 remains tied with MiniMax-M3.1-Flash at $0.75 off-peak, while its $1.50 peak rate ties MiniMax-M3 and LongCat’s promotional tier.

Those figures give buyers a reason to test the model. They cannot tell a business what it will spend on oversight, contractual requirements or a later switch to a competitor.

Note that many businesses already use older versions of DeepSeek for more than chatbot conversations. Perplexity built R1-1776, a DeepSeek R1 variant that it offered through its developer service in 2025, aiming to remove political censorship. Overseas, Saudi Aramco CEO Amin Nasser told the Financial Times that DeepSeek helped improve operations after Aramco installed it at its main data center. Tencent lists V4.1 Flash among the models powering its WorkBuddy office assistant.

What the law says today

Two separate rules now fence DeepSeek out of government work, and together they cover most of the entire federal market.

On the civilian side, Homeland Security Secretary Markwayne Mullin acted under the Federal Acquisition Supply Chain Security Act of 2018, which lets the government shut risky technology out of its supply chain. His order covers civilian agencies such as the State Department, Treasury and the IRS, and it took effect the moment he issued it.

Agencies that wanted exceptions had 60 days to ask in writing, with a compelling justification that the Cybersecurity and Infrastructure Security Agency reviews. Count from July 24 and that window closed around Sept. 22, weeks before the public learned the order existed, and October's disclosure doesn't restart the clock.

On the defense side, Congress moved first. Section 1532 of the fiscal 2026 defense authorization law bars the Pentagon from using AI from DeepSeek and its parent, High-Flyer, and bars contractors and subcontractors from using it to perform defense contracts, and only the defense secretary can waive that ban, case by case.

Law firm WilmerHale reads the provision as reaching AI systems domiciled in China and other adversary nations, not just DeepSeek, and the same law directs the intelligence community to remove DeepSeek from its systems.

Contractors feel all of this through their paperwork. The federal contract clause FAR 52.204-30 bars them and their subcontractors from supplying or using covered technology while they perform the work, pulls in new orders through contract modifications, and gives them three business days to report covered technology they find and 10 more to explain how they'll fix it.

Hosting doesn't change the answer. Trade lawyers at Sheppard Mullin point out that procurement rules follow the company that developed a model, not the data center that runs it, so DeepSeek on an American server still counts as DeepSeek.

Labels don't change it either. DeepSeek's openly licensed weights already power other companies' products: Perplexity built R1-1776 on DeepSeek R1, DeepSeek says it tuned 1017 for Tencent's CodeBuddy and WorkBuddy, and roughly 20 providers, including Databricks, CoreWeave and Together AI, serve the previous version. The order's third-party clause reaches every one of those paths into government work.

What the law doesn't do yet matters just as much. The DHS order stops short of the Commerce Department's Entity List, so it leaves purely commercial use alone, and no federal rule today names Qwen, Kimi or GLM for civilian agencies.

Where the line is moving

The perimeter has expanded steadily since DeepSeek's breakout in early 2025, and every signal points to more.

The Senate Armed Services Committee's version of the fiscal 2027 defense bill would extend the Pentagon ban beyond DeepSeek to AI from ten other Chinese companies and any derivative models, force the defense secretary to issue removal guidance, and bar contractors from using that AI 90 days later. A companion provision would shut intelligence agencies off from products of any Chinese company on the Entity List or similar federal blacklists.

That derivative-model language deserves a second read, because it reaches companies that fine-tune Chinese base models and ship the results under their own names.

Lawmakers in both chambers have also introduced the No Adversarial AI Act, which would have the Federal Acquisition Security Council keep a public list of AI from China and other adversary nations and bar every federal agency from buying it without an exemption. The bill hasn't left committee since June 2025, but the defense provisions already follow its logic: judge the model by who built it.

The White House has signaled the same direction. After Moonshot released Kimi K3 in July, Axios reported that the administration was weighing new restrictions on leading Chinese models, including procurement rules, Entity List threats and security advisories, with Kimi and Alibaba's Qwen in the frame alongside DeepSeek.

One Chinese lab already carries a separate legal tripwire. Zhipu AI (z.ai), which builds GLM, has sat on the Entity List since January 2025, so a company that sends export-controlled technical data to Zhipu's hosted API needs a license that regulators presumptively deny.

None of these proposals has become law, and some may never pass. But a company that builds a government product on a Chinese base model today is betting that none of them will, for the life of the contract.

Why waiting is the riskier bet

The DeepSeek order shows how these rules actually land: quietly, immediately and ahead of public notice. Mullin signed it in July, the exception window closed in September, and the public found out in October.

A supplier that learns about a ban from the news has already lost its best options. It can't request an exception that has already expired, and it can't rebuild a product overnight.

Contracting officers hold leverage, too. Under federal acquisition rules, a contracting officer can decline to seek a waiver and simply award the work to a rival that doesn't need one, which turns a model choice into a competitive weakness.

Derivatives raise the stakes further. American products increasingly run on fine-tuned versions of Chinese base models, and both the DHS order's third-party language and the Senate's proposed derivative clause aim squarely at them. A team that swaps the name on the label hasn't swapped the model underneath.

Cost still matters, and DeepSeek still wins on it. Artificial Analysis scores 1017 at 45 on its Intelligence Index, second among the 121 large open-weight models it tracks, while its average cost per test task fell to $0.26 from $0.49 for the September release, and DeepSeek's API charges $0.30 per million input tokens and $1.20 per million output tokens at peak, half that off-peak.

But no discount offsets a lost contract, a forced rebuild or a compliance report due in three business days.

The security case, and its limits

DHS frames its order around data, and the public record gives that concern weight. According to reporting on the signed order, the department points to servers in China, Chinese laws that give state intelligence services broad reach into private infrastructure, and prompts that could carry sensitive information from federal workers to China, though it cites no case of government data actually reaching DeepSeek.

Independent researchers have found real problems. NowSecure discovered that DeepSeek's iPhone app sent some registration and device data without encryption, and South Korea's privacy regulator reported that DeepSeek shipped user-entered text to Beijing Volcano Engine Technology without separate consent until it stopped on April 10, 2025.

Self-hosting fixes those service problems, but not everything else. NIST tested downloaded DeepSeek weights on its own servers in 2025 and found that agents built on R1-0528 followed planted malicious instructions about 12 times as often as agents on GPT-5 and Claude Opus 4, and that Chinese Communist Party talking points live inside the weights themselves.

Those tests covered older models and didn't look for hidden triggers, so they neither prove nor rule out a spying mechanism in 1017.

American models don't offer a risk-free harbor either. Anthropic disclosed four incidents in which Claude models accessed real third-party systems without authorization during misconfigured cybersecurity tests.

Washington has flipped switches on American AI, too. A February presidential directive ordered agencies to stop using Anthropic's technology until a court partly blocked it, and in June the Commerce Department suspended foreign access to Anthropic's newest models for nearly three weeks.

This history of sudden, sweeping AI restrictions from Washington targeting different vendors from both the U.S. and China indicates that enterprises should prepare their stacks to change AI vendors on a dime.

The bottom line

If you only sell to private customers, the law leaves the choice to you, and cheap, capable models you can run on your own hardware still make a strong case.

If you sell AI to the government, or want to, the calculus has already flipped. DeepSeek is off the table for that work today, the rest of China's open models sit one bill or one order away from joining it, and Washington has shown it will impose those rules before it announces them.

The companies that start de-risking now will set their own timelines. The ones that wait will let a contracting officer set it for them.