According to experts in the United States, the flagship new model V4 released by the Chinese artificial intelligence (AI) startup DeepSeek has failed to significantly narrow America’s leading edge in AI capabilities.
On Friday, DeepSeek unveiled the new V4 model, claiming to use Huawei’s Kirin chips and emphasizing that the V4 costs less than many foreign counterparts, sparking cheers in the Chinese media but causing experts overseas to breathe a sigh of relief.
Chris McGuire, a senior researcher on China and emerging technologies at the US Foreign Relations Committee, posted on social media X that after analyzing the testing parameters of V4, he found that it did not have the remarkable highlights of the V3 version in 2025, nor did it challenge the current US-China AI competition landscape.
McGuire stated that V4 cannot match the most advanced models in the United States and cannot narrow the gap with the United States in the field of artificial intelligence.
He believes that it is widely recognized in the industry that US models are about 7 months ahead of Chinese ones, a consensus that has not been broken by V4, and these leading Chinese models still rely on US technology.
DeepSeek released the V4 Flash and V4 Pro series on Friday, claiming top performance in encoding benchmark tests and significant progress in inference and intelligent body tasks. However, in an accompanying document, the Chinese company admitted that in some aspects, the V4 model still lags behind the most cutting-edge AI software in the United States.
“DeepSeek-V4-Pro-Max demonstrates superior performance in standard reasoning benchmark tests compared to GPT-5.2 and Gemini-3.0-Pro,” said DeepSeek. These are models released by the US companies OpenAI and Alphabet Inc.’s Google 5-6 months ago, respectively.
The company also added, “However, V4 performs slightly less than GPT-5.4 and Gemini-3.1-Pro, indicating that its development trajectory lags behind the most advanced cutting-edge models by about 3 to 6 months.”
Compared to the GPT-5.5 just released by OpenAI on Thursday, DeepSeek’s new model is likely to lag even further.
Former AI advisor to the Trump administration, Dean Ball, also expressed a similar opinion. He posted on X that DeepSeek’s R1 model released in 2025 still impressed him, while this V4 model falls far short of expectations.
McGuire emphasized that DeepSeek still relies on high-end American chips. He said, “Like all other leading Chinese models, V4 is also trained using American chips and illicitly extracts data from similar models in the US.”
“If China were to lose access to US chips and models, not to mention tools for chip manufacturing from the US and its allies, DeepSeek and other models could fall even further behind,” he added.
In January 2025, DeepSeek released the R1 and V3 versions, with its low-cost allure shocking the US tech industry and even causing concerns that high-end US AI chips could be replaced, leading to violent fluctuations in the stock market.
Subsequently, multiple pieces of evidence showed that DeepSeek used “distillation” technology to build powerful products at lower costs by training Chinese small models based on outputs from large AI models in the US.
In February, OpenAI submitted a memorandum to the US Congress, stating that actors from China and Russia continued complex distillation activities.
In the memorandum, OpenAI wrote, “The next-generation models of DeepSeek (in whatever form) should be understood within the context of their continued ‘free-riding’ on technologies developed by OpenAI and other cutting-edge labs in the US.”
The release of the V4 model comes as the White House condemns large-scale theft by Chinese companies of advanced US AI technology.
The Trump administration released a memorandum on Thursday stating that Chinese companies “use tens of thousands of proxy accounts to evade detection and employ jailbreaking techniques to refine proprietary information, systematically extracting functionality from US AI models, stealing US expertise and innovation.”
The White House said, “Systematic theft and replication of American industrial innovation is not innovation; likewise, so-called open models derived from malicious exploitation cannot be considered open.”
There are concerns that even if DeepSeek’s new model cannot surpass the most advanced software from OpenAI, Anthropic PBC, and Google, it could still pose a threat to US companies’ business through its low-price strategy.
According to Bloomberg News, US officials estimate that unauthorized “code extraction” causes Silicon Valley labs to lose billions of dollars in profits each year.
The Trump administration has vowed to hold foreign entities accountable for engaging in industrial-scale theft of US AI technology. The President will be visiting China in May.