The White House is reportedly considering a rule banning installing the DeepSeek app on government devices.
Based on anonymous sources, the Wall Street Journal reports that the measure is likely to be implemented. According to the report, there are concerns about the way DeepSeek processes user data. The Chinese AI lab does not disclose details about who has access to the information collected.
DeepSeek gained notoriety earlier this year with the launch of DeepSeek-R1, an open-source large language model optimized for reasoning. This model outperforms OpenAI’s competing o1 reasoning algorithm on several tasks. Moreover, DeepSeek claims that R1 has cost less to train than many previous LLMs.
DeepSeek chatbot app is popular
In addition to R1, DeepSeek offers a chatbot app similar to ChatGPT, explains SiliconANGLE. This service is the target of the U.S. government’s possible ban. In fact, at one point, DeepSeek was the most downloaded app in both the App Store and Google Play.
DeepSeek’s mobile app is not based on R1 but on DeepSeek-V3, an LLM that the AI Lab made available as open source in December. This model has more limited reasoning capabilities. However, since R1 is based on V3, it has many architectural similarities with R1.
Both models contain 671 billion parameters, organized into subnetwork neural networks that each focus on specific tasks. When a user enters a prompt, only one subnetwork generates the response. This limits hardware usage.
Generating multiple tokens simultaneously
DeepSeek trained V3 with a dataset of 14.8 billion tokens. A token can consist of a few letters or numbers. Usually, LLMs generate output tokens by token, but with V3, DeepSeek took a different approach: the model was configured to generate multiple tokens simultaneously, which the company says improved performance.
R1, DeepSeek’s most advanced reasoning model, is a version of V3 trained more intensively. This additional training included partly supervised fine-tuning, where the model was given examples of performing tasks. In addition, DeepSeek further refined R1 with reinforcement learning.
According to The Wall Street Journal, the White House’s possible DeepSeek ban would not only be limited to government devices. One would also consider banning app stores from offering the chatbot service. In addition, there are discussions about restrictions on how U.S. cloud providers may provide DeepSeek models to customers. Discussions on the latter two measures are still in the early stages.
It is unclear whether the restrictions on cloud providers affect only R1 and V3 or DeepSeek’s less powerful LLMs.
In January, the company released a reasoning model called R1-Zero, which was trained entirely with reinforcement learning. Typically, LLM developers combine this with supervised fine-tuning. DeepSeek claims that R1-Zero is the first open source model demonstrating that LLMs’ reasoning capabilities can be improved purely by reinforcement learning.
In addition, DeepSeek has released several so-called distilled models based on R1 as open source. These models contain some of R1’s knowledge and are based on the open-source Llama and Qwen-LLM families. Their size ranges from 1.5 billion to 70 billion parameters.
The U.S. Navy and NASA have already prohibited their personnel from installing the DeepSeek app on work devices. Texas, New York and Virginia recently introduced similar rules for state employees. In South Korea and Italy, privacy regulators have banned appstores from offering DeepSeek to consumers.