Moonshot announces retirement of first-generation multimodal model kimi K2.5

Moonshot, a Chinese artificial intelligence company, has announced that its first-generation multimodal model, Kimi K2.5, will reach the end of its service life by the end of the month. The company previously released Kimi K2.5 in January of this year, along with the launch of the K2.5 version of its intelligent assistant. Users on the official website automatically transitioned to the K2.5 model in the chat interface. Moonshot described Kimi K2.5 as its most intelligent model to date, achieving state-of-the-art performance in areas such as agents, code, images, video, and various general intelligence tasks. It is also the most versatile model in the Kimi series, featuring a native multimodal architecture that supports both visual and textual input, as well as thinking and non-thinking modes, and conversation and agent tasks. In July, Moonshot officially open-sourced the Kimi K3 model, a 2.8 trillion parameter model built using the Kimi Delta Attention mechanism and Attention Residuals technology. The model natively supports visual understanding and has a 1 million token context window. Moonshot stated that Kimi K3 is the first open-source model to reach the 2.8 trillion parameter scale, and during the past 12 months, Kimi models have consistently set the upper limit for open-source model sizes.