Chinese startup "DeepSeek" has launched its new AI model DeepSeek-V4, which supports command inputs as long as 1 million words, while significantly reducing processing and memory costs amid intensifying technology competition between China and the United States.
"DeepSeek," a rising startup based in Hangzhou, China, known for the "DeepSeek Shock" that caused global tech stocks to plummet last year, announced its latest artificial intelligence model "DeepSeek-V4," stating that this new version has been developed to offer higher capabilities while drastically reducing resource usage.
The company stated via announcements on WeChat and X platforms that DeepSeek-V4 supports an exceptionally long maximum input length of up to 1 million words, which is a world-leading level among both national and open-source models, enabling the AI to process vast amounts of data and accurately understand complex commands.
For this launch, DeepSeek divided the model into two sub-versions to suit different use cases: DeepSeek-V4-Pro, the flagship version with 1.6 trillion parameters, boasting global-level intelligence only slightly behind Google's Gemini-Pro-3.1 but surpassing all open-source models; and DeepSeek-V4-Flash, a cost-effective, efficient, and fast version with 284 billion parameters, designed as an economical option for businesses and developers.
Zhang Yi, founder of research firm iiMedia, commented that this launch marks a "major turning point" for the industry because it addresses the issues of slowing performance and excessively high costs when processing long commands, potentially moving advanced AI technology from a research tool to widespread commercial use.
The DeepSeek-V4 launch comes amid a period of intense tension between the US and China, with the White House accusing Chinese agencies of a major effort to steal American AI technology.
Michael Kratsios, science and technology advisor to former President Donald Trump, stated that the US has evidence of Chinese operations using "Distillation" techniques—extracting data from competitors' models at an industrial scale to create their own cheaper and similarly efficient models.
DeepSeek's success in creating high-performance AI at low cost has posed a difficult challenge for Silicon Valley companies like Meta and Microsoft, which must invest heavily to maintain competitiveness; recently, Meta announced plans to lay off another 10% of its workforce to preserve profits and focus funding on AI development.
However, despite DeepSeek's high popularity in China, especially in healthcare, finance, and local government sectors, it faces concerns regarding data privacy and content censorship—issues common to Chinese AI, which often avoids answering politically sensitive questions.
. AFP