Moonshot AI just released Kimi K3, a 2.8 trillion-parameter open-weight model that immediately slots into conversations previously reserved for the biggest closed systems from US labs. The Beijing-based startup launched the model on July 16 at the World Artificial Intelligence Conference in Shanghai, where Chinese President Xi Jinping used the occasion to call for international cooperation on AI development.
Kimi K3 is the first open model to enter what researchers are calling the 3-trillion parameter class, and early benchmarks suggest it can hang with the likes of Claude Fable 5 and GPT-5.6 Sol.
What makes Kimi K3 different
The model uses a Mixture-of-Experts (MoE) architecture. Instead of running all 2.8 trillion parameters on every query, the system activates only a subset of its 896 experts depending on the task. It also features something Moonshot is calling “Kimi Delta Attention” and “Attention Residuals,” proprietary techniques that appear to improve how the model handles long sequences of information.
Kimi K3 supports a 1 million-token context window. Moonshot describes its target use cases as “long-horizon coding” and knowledge work productivity. The model also ships with native multimodal capabilities, including 3D generation. On LMArena’s Frontend Code Arena benchmark, Kimi K3 debuted at the number one spot, with a third-place overall ranking across broader evaluations.
