In May, Google was supposed to release the next version of its flagship AI model, Gemini 3.5. It did not. According to Bloomberg, the company is still working to improve the model's coding abilities — a direct attempt to close the gap on Anthropic's Claude, which is widely regarded as the most capable AI coding tool available.

OpenAI, meanwhile, had already released a new model at the start of July. When Google's delay became public, its share price fell, as investors began to question whether the world's most prominent search company could hold its position at the front of the AI race. The timing could hardly have been worse. On the same week that Google's struggles were laid bare, a Chinese startup called Moonshot quietly released Kimi K3 — and the AI world took notice.

The model demonstrated surprisingly strong coding skills, particularly in front-end development, and spread so rapidly that Moonshot's servers struggled to cope. "Kimi K3 has received far more love than we expected, and our GPUs are feeling it," the company posted on X, noting that demand had pushed close to the limits of its capacity within 48 hours. Anastasios Angelopoulos, the CEO of AI testing company Arena, called it "the single biggest release of the year." What makes Kimi K3 particularly significant is not just its performance, but its philosophy.

Unlike ChatGPT, Claude or Gemini, it is free to use and open-weight, meaning that anyone can download its core components and run it on their own computer. This approach — releasing powerful models openly and at low cost — has become a defining strategy for Chinese AI developers.