Chinese AI startup Moonshot AI has unveiled its next-generation flagship model, Kimi K3. It supports a 1 million token context window and can handle text, images, and video simultaneously. The model is designed for long-form coding, research, and knowledge work.
K3 boasts a total of 2.8 trillion parameters using a mixture-of-experts (MoE) architecture, activating only 16 out of 896 experts per inference. Moonshot AI claims it's the first open model at the 3-trillion-parameter scale.
According to official benchmarks, K3 outperforms Claude Opus 4.8 and GPT-5.5 on some coding and agent tests, though it still lags behind Claude Fable 5 and GPT-5.6 overall.