Hangzhou-based DeepSeek has released DeepSeek-V4, its first major new artificial intelligence model in over a year, positioning the Chinese startup once again at the center of a deepening rivalry between Beijing and Washington over technological supremacy.
The release comes as the White House has accused Chinese entities of running industrial-scale campaigns to steal American AI technology, a claim Beijing has called “baseless.”
DeepSeek first burst onto the global scene in January 2025 with its R1 reasoning model, a large language model that performed at a level comparable to ChatGPT and other leading United States chatbots.
The company claimed it had developed the model at a fraction of the cost of its Western rivals, triggering a sharp sell-off in AI-related stocks on Wall Street.
A man takes photos of a DeepSeek display at a shopping mall in Hangzhou, Zhejiang province, China, April 23, 2026. (AFP Photo)
New model raises the bar
DeepSeek-V4 is available in two versions: DeepSeek-V4-Pro and DeepSeek-V4-Flash. The latter, according to the company, is a “more efficient and economical choice” due to its smaller footprint. V4-Pro carries 1.6 trillion parameters, while V4-Flash carries 284 billion.
The model supports a context length of one million tokens, on par with Google’s Gemini, and has been described by the company as “world-leading” in reducing compute and memory costs.
Zhang Yi, founder of tech research firm iiMedia, called the release “a genuine inflection point for the industry,” saying ultra-long context support could move long-text processing out of high-end research labs and into mainstream commercial use.
An aerial view of the building housing the headquarters of Chinese AI startup DeepSeek is seen in Hangzhou, Zhejiang province, China, Jan. 14, 2026. (AFP Photo)
Censorship concerns and bans
Despite its technical achievements, DeepSeek’s tools have faced significant scrutiny outside China. Like other Chinese AI platforms, DeepSeek avoids topics typically censored in China, including the 1989 Tiananmen crackdown.
Those restrictions, combined with data privacy concerns, have led to the platform being banned or restricted on government-issued devices in the United States, Australia, and South Korea.
DeepSeek currently holds 4% of the global chatbot market share, according to web traffic analysis firm Similarweb. ChatGPT dominates the sector at 68 percent.
The DeepSeek AI application logo is displayed on a smartphone screen beside the letters “AI” on a laptop screen in Frankfurt am Main, Germany, Sept. 1, 2025. (AFP Photo)
Open-source edge
One factor behind DeepSeek’s rapid adoption, particularly in developing countries, is its open-source architecture.
Unlike the proprietary models sold by OpenAI and other Western rivals, DeepSeek’s systems make their inner workings publicly available, allowing programmers to customize the software for their own needs.
The Chinese government has highlighted this approach as a competitive advantage. “Chinese AI models are leading the way in the open-source innovation ecosystem,” National People’s Congress spokesman Lou Qinjian told policymakers.
Web pages of the Chinese AI platform DeepSeek are displayed on screens in Goyang, South Korea, Feb. 17, 2025. (AFP Photo)
Chip smuggling allegations
DeepSeek’s rise has not been without controversy. According to reports, including one by the technology website The Information, the company allegedly used thousands of chips that were dismantled in third countries and smuggled into China to train its new V4 model, thereby evading the US ban on the export of advanced chips to China.
DeepSeek did not respond to requests for comment. NVIDIA, whose chips were cited in the reports, told The Information it had seen no evidence of such activity and called the smuggling claims “far-fetched.”
On the day of V4’s release, DeepSeek said the model was trained using chips from both Nvidia and Huawei’s domestically produced Ascend line.
Huawei, sanctioned by the United States since 2019, confirmed its full range of Ascend products supports the V4 series.
April 24, 2026 05:21 PM GMT+03:00
