Chinese AI lab DeepSeek on Friday released two preview versions of its newest large language model, DeepSeek V4, positioning the release against competing closed-source platforms from OpenAI, Anthropic, and Google. The close collaboration with Huawei on the new model contrasts sharply with DeepSeek's past reliance on Nvidia chips. DeepSeek did not disclose which processors it used to train the latest model. The launch lands one day after White House science and technology adviser Michael Kratsios issued a formal memo accusing Chinese entities of running coordinated campaigns to extract capabilities from American-built AI systems, sharpening the regulatory backdrop against which the release landed. [1][2]
DeepSeek released two versions of the model: the V4-Pro, which carries 1.6 trillion parameters, making it the company's largest model by that metric, and the smaller V4-Flash, which has 284 billion parameters. DeepSeek says both versions are open-source models with significant improvements in knowledge, reasoning, and "agentic" capabilities, meaning the ability to perform complex tasks and workflows autonomously. Both versions carry a 1 million token context window, a substantial expansion from the 128,000 token window supported by DeepSeek's prior V3 model.[3] Huawei confirmed in a separate statement Friday that its Ascend chips and related technology are compatible with the V4 models.[4]
The chip dimension carries the most immediate policy weight. Washington began restricting China's access to advanced AI chips produced by U.S. firms in 2022, and Beijing has since accelerated its push to achieve technology self-sufficiency. Huawei, whose Ascend AI chip line is central to China's effort to reduce dependence on U.S. semiconductor technology, said it had worked closely with DeepSeek so the V4 models could run across its full line of high-performance systems.[4] DeepSeek has acknowledged prior use of Nvidia chips but has not stated whether those chips were subject to U.S. export controls.[1] The Trump administration in January approved sales of Nvidia's H200 chips to China, but shipments have been delayed by disagreements over sales terms on both sides.[1]
The intellectual property dimension of the V4 launch is legally unresolved but operationally significant. In a Thursday memo, Michael Kratsios, the president's chief science and technology adviser, accused foreign entities "principally based in China" of engaging in deliberate, industrial-scale campaigns to extract capabilities from leading U.S.-built AI systems.[5][6] In February, Anthropic accused DeepSeek, Moonshot AI, and MiniMax of running campaigns to illicitly extract capabilities from its Claude chatbot, characterizing the conduct as intellectual property theft.[7] OpenAI, in a February letter to U.S. lawmakers, made similar allegations, arguing that China should not be permitted to advance what it called "autocratic AI" by appropriating American innovation.[5] DeepSeek has denied intentionally using synthetic data generated by OpenAI and has not directly addressed the Anthropic allegations in connection with V4.[1] No formal litigation has been filed as of publication, and neither Anthropic nor OpenAI has indicated it is pursuing civil claims at this stage.
The V4 release fits into a broader arc of U.S.-China technology decoupling that has accelerated since the Export Administration Regulations were amended in October 2022 to restrict shipments of advanced chips to China. DeepSeek's pivot to Huawei hardware represents a concrete test of whether Chinese AI development can achieve competitive performance outside the Nvidia ecosystem. According to a recent report from Stanford University's Institute for Human-Centered AI, the performance gap between top U.S. and Chinese AI models has "effectively closed."[6] DeepSeek, owned by China's High-Flyer Capital Management, is reportedly seeking to raise funds at a valuation exceeding $20 billion, with Alibaba and Tencent said to be in discussions to acquire stakes.[1] Whether Washington treats the Huawei partnership as a sanctions or export-control matter, in addition to the existing IP-theft framing, will determine the next pressure point in this dispute.
References:
[1] Bloomberg/BNN Bloomberg. (2026, April 24). DeepSeek previews new AI model adapted to run on Huawei chips. https://www.bnnbloomberg.ca/business/artificial-intelligence/2026/04/24/deepseek-previews-new-ai-model-adapted-to-run-on-huawei-chips/
Sources:
Comments (0)