Why DeepSeek’s new model was snubbed

The latest version of DeepSeek reaches the same heights as its predecessor. According to tests conducted by the company, its most powerful “Pro” system performs only slightly worse than models launched by leading U.S. rivals three to six months ago. DeepSeek v4 is also very cheap for customers. The introductory offer puts it at a thousandth of the price of the best American models for certain purposes. Even after that rate expires on May 7, the v4 will cost one-tenth to one-quarter of its U.S. counterpart.

But unlike DeepSeek’s previous bombshells, v4 doesn’t appear to be cheap to build. In 2025, the lab is keen to point out, training its AI will cost around $6 million, well below current levels in the West. The lab’s technical white paper on v4 omits any estimate of this measure. The fact that there was a 16-month gap between v4 and its predecessor also suggests that it used a lot of processing power to train it.

The launch comes as China’s artificial intelligence field becomes increasingly crowded. DeepSeek faces increasing competition from other independent labs such as Moonshot and Z.ai, as well as Chinese internet giants. The Qwen series, made by e-commerce giant Alibaba, has been at the top of China’s charts for much of the past year. ByteDanceThe creator of TikTok and the maker of Doubao, China’s most popular chatbot. Dola (known as Dola outside of China) is extremely popular in Mexico, the Philippines and the UK, ranking higher than Google’s Gemini in the Apple App Store.

In China, much of the attention has shifted to AI-based applications. Alibaba is applying its Qwen model to other areas of its business, such as providing a “digital workforce” to merchants using its e-commerce platform. The country’s internet giants are racing to build AI-powered “super apps” Can facilitate a wide range of digital transactions. Merely smart models are not seen as a way to make money from technology.

At the same time, DeepSeek has had to contend with greater state intervention. The Chinese government has been promoting chips made by national semiconductor champion Huawei. DeepSeek reportedly tried to train new models on them, but ended up switching to Nvidia’s chips, adding cost and time. It seems unlikely that the government will give local AI companies a free rein anytime soon: On April 27, the government said it would block US social media giant Meta from acquiring Manus, another of the country’s AI darlings.

DeepSeek’s latest version fails to dazzle, and that’s nothing to lament. US lab Anthropic recently decided that its leading Mythos model was too powerful to be released to the public due to its hacking capabilities. In contrast, the documentation that comes with DeepSeek v4 makes no mention of security measures at all. If Chinese labs do catch up to their American counterparts, they may not show the same restraint.