DeepSeek Enhances AI Model V3

Chinese AI startup DeepSeek has released a major update for its language model V3. DeepSeek-V3-0324 is available on the Hugging Face platform.

The new version of the neural network has been improved in areas of reasoning, programming, and writing. It surpasses its predecessor and solutions from competing companies in several benchmarks.

DeepSeek became a notable player in the artificial intelligence market in January. At that time, the company released an open-source reasoning-oriented model, DeepSeek-R1, developed on a significantly smaller budget.

V3, with 671 billion parameters, emerged in December 2024.

CEO of the Chinese startup 01.AI, Lee Kai-fu, believes that China has narrowed the gap with the US in AI development to three months, as companies like DeepSeek have learned to use chips and apply algorithms more efficiently.

The country has already become a leader in software development for infrastructure, he added.

“Previously, I think the gap was six to nine months, and we were lagging in everything. Now I think we might be three months behind in some core technologies, while in other specific areas we are ahead,” said Kai-fu.

US sanctions created short-term problems but also forced Chinese companies to innovate, the expert emphasized.

Following the rapid rise in popularity of DeepSeek’s AI models, the company caught the attention of Chinese authorities. Employees are working under new, stricter conditions.

Previously, OpenAI recommended that the US government ban the Chinese competitor’s solutions, as the project is “state-subsidized” and “state-controlled.”