U.S. Bans DeepSeek From Federal Civil Systems as Chinese AI Firms Push New Models

The notice, published in the Federal Register, applies to all civil departments’ IT systems, the reports said. DHS required removal of all DeepSeek products, all development results linked to the company and third-party derivative products built with DeepSeek outputs. Covered federal civil agencies are also prohibited from buying any DeepSeek-related products in the future. The U.S. side said DeepSeek’s development outputs constitute a threat.

DeepSeek did not issue a public response in the reports. The ban came as DeepSeek founder Liang Wenfeng was increasing his exposure to physical AI. According to media reports, former Tencent AI Lab and LimX Dynamics humanoid robot foundation model head Yu Tao has started a company focused on LLM-plus-embodied AI. The company’s first funding round valued it at $70 million and a second at $200 million, with DeepSeek among investors, people cited by media said.

Days earlier, reports said DeepSeek planned to lead a funding round for Winverse AI, a robotics AI startup founded by computer vision scientist Hu Han, who recently left Tencent. The round was reported at tens of millions of dollars and a post-money valuation of about $400 million. In August, DeepSeek invested 140.8 million yuan in Unitree Robotics’ IPO strategic placement and signed a cooperation agreement with Unitree. Multiple sources said Liang has been meeting embodied AI entrepreneurs and weighing both building a team and investing.

In the model market, MiniMax released M3.1 Flash on Oct. 19. The company said the Flash model handled more than 200 trillion tokens anonymously under the name Space Bunny before its release. MiniMax said M3.1 Flash improves on agentic tasks, coding, computer use and cybersecurity, with online inference nearly three times faster and usage costs 70% lower. It uses a more aggressive sparse attention architecture, with daily output speed above 200 tokens per second and a median of more than 170 tokens per second. API prices are 1 yuan per million input tokens, 4 yuan per million output tokens and 0.1 yuan per million cached-read tokens. A Pro version is due later, MiniMax said.

ByteDance’s Seed model evaluation head Huang Wenhao left recently, according to people familiar with the matter cited by media. Some reports said he plans to start a business, but the direction and financing were not known. Huang joined ByteDance’s Seed team in August 2024 as a technical project management and planning head after serving as a co-founder, algorithm vice president and model training head at 01.AI. He reported to then-Seed head Zhu Wenjia. After Google DeepMind vice president Wu Yonghui took over Seed in early 2025, Huang reported to Wu and later focused on Seed model evaluation, frontier technology planning discussions and recruiting.