Kimi K3: The Chinese AI model challenging America's lead

Kimi K3: The Chinese AI model challenging America's lead

Can a Chinese AI model with 2.8 trillion parameters and laughably low costs put America's tech giants on the back foot? The launch of Kimi K3, an open-weight model from the Chinese company Kimi, has peine the debate over who is really winning the AI race. While the West spends billions on datacenters, Beijing is betting on lightweight, cheap, open-source models that anyone can run.

Groundbreaking technical specifications

Kimi K3 is no ordinary model. With 2.8 trillion parameters, 16 active experts per token and a 1 million token context window, it is optimized for long tasks and autonomous agents. What is revolutionary is that, being open-weight, it can run on consumer hardware: it has been shown to work on a simple Mac Mini, replicating a desktop environment in 3 hours. "For a reasonable price, a vast number of small companies can build their own rigs without depending on third parties," the technical analyses explain. The cost per inference is a fraction of that of American models like GPT-4 or Claude.

Open warfare between China and the US

The launch of Kimi K3 comes amid escalating tensions. Anthropic, the maker of Claude, alleged that Alibaba used around 25,000 fraudulent accounts to extract data from the American model, generating some 29 million exchanges. The Chinese strategy seems clear: offer models of similar quality at lower costs to erode the profitability of American datacenters. "Your datacenters will never be profitable, we're going to ruin you," experts summarize. Still, some point out that Kimi K3 is actually a clone of models like Fable and Sol, and that its inference estimulante ilegal remains low.

Europe, absent from the race

While China and the US fight for leadership, the European Union is conspicuously absent. The number of AI models with European parent companies is zero. "In 15 years we have gone from 25% of global GDP to 17%," analysts lament. Regulation (the AI Act) and a lack of structured funding have weighed down any initiative. The irony is that while Europe debates whether AI has gender bias or how to optimize bottle caps, the Chinese have already hacked Mac Minis and are turning consumer graphics cards into server GPUs.

Artificial intelligence has become a geopolitical weapon. Whoever controls cheap, accessible models will control the future. And for now, China is several steps ahead. Europe, meanwhile, is still debating the design of bottle caps.