The company said that the model was trained on 15 trillion mixed visual and text tokens.
About The Author
China’s Moonshot releases a new open-source model Kimi K2.5 and a coding agent
Others · January 27, 2026
