U.S. Bars DeepSeek From Federal Systems, Firm Open-Sources AI

The U.S. government ordered federal civilian agencies to remove DeepSeek products from their information systems and prohibited future purchases of third-party services built with the Chinese company’s artificial intelligence models.

The Department of Homeland Security published the directive on Oct. 20. Secretary Markwayne Mullin signed it on July 24 under the Federal Acquisition Supply Chain Security Act of 2018. Agencies must begin complying immediately and have 60 days to request an exception.

The order covers DeepSeek software and services, along with products from other companies that use DeepSeek implementations. Federal contractors must check whether their products use the company’s models. They may need to replace those components or seek approval to continue serving government customers.

DHS cited the risk that users could enter sensitive information into DeepSeek services and that the data could be transmitted to servers in China. The department also warned that the technology could contain security weaknesses and that information collected through the services could help reveal details about U.S. government activities.

The restrictions follow earlier actions against DeepSeek by the Commerce Department, the U.S. Navy and several state governments. China condemned the order, accusing Washington of politicizing technology and harming commercial cooperation between the two countries.

On Oct. 20, DeepSeek announced the release of DeepInfer, the inference framework it uses to operate large-scale AI services. The company also released deepseek-recipe, a set of tools for token encoding, prompt formatting and application programming interface request processing.

DeepInfer includes computing kernels, distributed inference functions, speculative decoding and other processing tools. These components are designed to increase the number of responses a system can generate while reducing the computing resources required. They allow developers to reproduce parts of DeepSeek’s API behavior when running the company’s models.

In a statement, company engineer Cui Tianyi described DeepInfer, deepseek-recipe and other components as tools for serving DeepSeek models. The release is intended to allow businesses and organizations to deploy the technology and generate revenue from it, according to the statement.

The software is available under the Apache 2.0 license, which permits commercial use, modification and redistribution. DeepSeek’s release expands its open-source offerings beyond model weights by providing infrastructure that developers can run on their own systems.

DeepInfer currently supports DeepSeek-V4.1-Flash, a 552-billion-parameter model introduced in September. The model activates only part of its parameters for individual tasks. DeepSeek described its design as requiring less memory and storage for inference caching, allowing operators to process more requests with fewer computing resources.

The framework supports Nvidia SM100 graphics processors and Huawei Ascend 950 chips. The initial release provides offline inference, but operators must add systems for cache management and load balancing to create a complete production service.

Businesses and cloud providers outside the U.S. federal market can use the tools to run DeepSeek models on their own infrastructure. Private hosting does not exempt a supplier from the federal restrictions, which also cover third-party products that use DeepSeek implementations.