DHS orders DeepSeek removal from federal systems as DeepSeek open-sources DeepInfer

What the DHS directive covers

The directive doesn't stop at DeepSeek's own products. According to CryptoSlate's report, it also covers software and services from other companies that use DeepSeek implementations, which could bring American contractors whose products include DeepSeek models into scope.

Homeland Security Secretary Markwayne Mullin signed the order on July 24 under the Federal Acquisition Supply Chain Security Act of 2018, per the report. Agencies had to start complying immediately (with a 60-day window to request exceptions).

DHS said its concerns centered on DeepSeek's operations in China, including the possibility that sensitive information entered by government employees could be sent to servers there. The department also warned that DeepSeek's technology could contain security vulnerabilities and that information collected through its services could provide intelligence on US government activities.

It's not the first move. The order followed earlier actions against DeepSeek by the Commerce Department, the US Navy and several state governments. CryptoSlate reported that China condemned the decision, accusing Washington of politicizing technology and undermining commercial cooperation between the two economies.

Contractors face new compliance work

Federal technology suppliers now have to assess whether products using DeepSeek implementations can keep serving government customers. In CryptoSlate's assessment, affected contractors may need to replace DeepSeek components or seek exemptions, a shift that could benefit competing AI providers.

DeepSeek opens up its serving stack

DeepSeek went the other direction on the same day.

The company open-sourced DeepInfer, which includes optimized computing kernels, distributed inference, speculative decoding and processing tools aimed at improving throughput and cutting compute needs. It also released deepseek-recipe, a set of components that handle token encoding, prompt formatting and API request processing.

“We are releasing DeepInfer, deepseek-recipe, and other vital components for serving DeepSeek models,” company engineer Cui Tianyi said

Cui said the goal was to let more businesses and organizations worldwide deploy the technology and profit from it. The release takes DeepSeek's open-source strategy beyond publishing model weights.

DeepInfer ships under the Apache 2.0 license (which permits commercial modification and redistribution) and supports Nvidia SM100 graphics processors and Huawei Ascend 950 chips. According to DeepSeek, it currently supports DeepSeek-V4.1-Flash, a 552-billion-parameter model introduced in September that activates only a fraction of its parameters for each task. The company said that architecture needs much less memory and storage for inference caching, so providers can handle more requests with fewer computing resources.