China’s AI ambitions just got a lot more concrete. Moonshot AI, a Beijing-based startup founded in early 2023, has unveiled Kimi K3, a model packing 2.8 trillion parameters and a one-million-token context window. The company claims it outperforms some of the biggest names in American AI on key benchmarks.
What Kimi K3 actually brings to the table
Kimi K3 uses a mixture-of-experts architecture, which essentially means the model doesn’t activate all 2.8 trillion parameters for every query. This approach makes the model more efficient to run despite its massive size.
According to initial performance claims, Kimi K3 outperforms Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5 on benchmarks related to coding and long-horizon tasks. It still trails the absolute top-tier closed models like Claude Fable 5 and GPT-5.6 Sol, but the gap is apparently narrow enough to make Silicon Valley uncomfortable.
The model also features native multimodality, meaning it can handle text, images, and potentially other data types without bolting on separate systems.
