China’s DeepSeek AI programme developer astonishes the AI community

DeepSeek, a Chinese AI software programme developer, is causing consternation in the AI community by saying it can train an industry leading programme at a cost of $5.6 million compared to the $100 million to $1 billion cost cited by top US AI developers

Last week DeepSeek launched a programme called R1, for complex problem solving, that was trained on 2000 Nvidia GPUs compared to the 10s of thousands typically used by AI programme developers like OpenAI, Anthropic and Groq.

Besides R1, DeepSeek has a programme called V3. R1 and V3 together were rated in the top ten AI models on the University of California at Berkeley’s AI rating service, Chatbot Arena, beating Anthropic’s Claude and Grok from Elon Musk’s xAI.

“DeepSeek-R1 is now live and open source, rivalling OpenAI’s Model o1, available on web, app, and API,” says DeepSeek’s website, adding “V3 achieves a significant breakthrough in inference speed over previous models. It tops the leaderboard among open-source models and rivals the most advanced closed-source models globally.”

V3 is free but companies that want to hook up their own applications to DeepSeek’s model and computing infrastructure have to pay to do so.

It is designed for tasks like coding, mathematics, and reasoning. DeepSeek Coder uses neural networks to generate code in over 80 programming languages, using architectures like Transformer and Mixture-to-Expert.

The company is led by Liang Wenfeng, a former hedge fund manager who used AI techniques to run his fund called High-Flyer and then launched DeepSeek to spin off the AI technology.