Kimi K3 is the first open 3T-class model, with a 1M-token context window that ingests entire codebases and long documents without chunking. Its sparse expert architecture (2.8T parameters, 16 of 896 active per token) delivers frontier-scale capability with efficient inference, and native multimodal support lets it reason over text, images, and video together. Tuned for max thinking effort by default, it sustains long-horizon autonomous work like coding, research automation, and multi-hour agentic sessions.