The U.S. Department of Homeland Security on October 20 directed federal civilian agencies to remove DeepSeek products from their information systems and prohibited future purchases. The restrictions also cover software and services developed by third parties that use DeepSeek models, putting technology suppliers serving the government under pressure to review their products, identify affected components and assess replacement options or exemptions.
On the same day, DeepSeek opened access to its inference framework and supporting tools, giving developers more scope to deploy its models independently. The two developments affect separate but connected parts of the artificial intelligence market: the U.S. federal technology procurement chain and companies' choices for deploying AI systems.
Federal restrictions extend to third-party software
The directive was signed on July 24 by Homeland Security Secretary Markwayne Mullin under the 2018 Federal Acquisition Supply Chain Security Act. Agencies must begin implementing it immediately and may apply for exceptions within 60 days.
The order covers more than products supplied directly by DeepSeek. It also applies to software and services built by other companies on DeepSeek models, potentially affecting U.S. government contractors that have integrated the technology into their products. Vendors will need to determine whether those integrations fall within the scope of the directive and whether an exemption is available.
The department cited national security concerns, including the possibility that government employees could enter sensitive information into services that might transmit data to servers in China. It also warned that vulnerabilities in the technology and the data collected by the services could expose information about U.S. government activities. The directive does not automatically exempt third-party products operated on private infrastructure, leaving suppliers to assess their status under the order.
The measure follows earlier efforts by the U.S. government to restrict Chinese technology in government infrastructure. The Commerce Department, the U.S. Navy and several state governments have previously taken action involving DeepSeek. Chinese officials have criticized the decision, arguing that Washington is politicizing technology issues and damaging commercial cooperation between the two countries.
For government technology suppliers, the practical impact will depend on whether their products use DeepSeek technology and whether they can secure an exception. Some companies may have to replace affected components, while competing vendors could gain opportunities in government contracts.
DeepSeek releases inference framework and deployment tools
DeepSeek also announced the open-source release of DeepInfer, the inference framework that supports its large-scale AI services. Technology previously used internally is now available to developers, including optimized computing kernels, distributed inference capabilities, speculative decoding and processing tools intended to increase throughput while reducing the computing resources required to generate responses.
The company also released deepseek-recipe, which includes components for token encoding, prompt formatting and API requests. Developers can use the tools to reproduce key elements of DeepSeek's API behavior when running its models.
DeepSeek engineer Cui Tianyi said the release of DeepInfer, deepseek-recipe and other model-serving components was intended to allow more companies and organizations worldwide to deploy and use the technology.
The release goes beyond model weights by making some of the infrastructure needed to run the models available to developers. That could give businesses more control over their deployment environments and reduce reliance on proprietary AI services from OpenAI, Anthropic and Google. It does not, however, override the U.S. federal procurement directive: the order expressly covers third-party products built using DeepSeek technology.
DeepInfer supports multiple chips but lacks some production components
DeepInfer currently supports DeepSeek-V4.1-Flash, a model launched in September with 552 billion parameters. Only a portion of those parameters is activated for an individual task. DeepSeek said the architecture can substantially reduce the memory and storage required for inference caching, allowing service providers to handle more requests with fewer computing resources.
The framework is licensed under Apache 2.0, which permits commercial modification and redistribution. It supports Nvidia's SM100 graphics processors and Huawei's Ascend 950 chips, giving cloud providers and enterprises more options for deploying the model across different hardware platforms. Support for Huawei chips is also relevant to the development of China's domestic AI infrastructure.
The initial release provides offline inference rather than a complete production service. Operators still need to supply components such as cache management and load balancing before they can build a full-scale service. That distinction matters for companies weighing self-hosted deployment while also reviewing whether their use of DeepSeek-based software is compatible with the new federal procurement restrictions.