The Moonshot announced today via Kimi's official Weibo that its first trillion-parameter multimodal model, Kimi K2.5, will retire at the end of this month. The model was launched and open-sourced in January this year, and it is the most versatile model in Kimi's history so far, featuring a native multimodal architecture that supports both visual and text inputs, thinking and non-thinking modes, as well as dialogue and Agent tasks.
K3 Takes Over, Parameter Scale Rises Again
K2.5 achieved state-of-the-art performance in open-source models on tasks such as Agents, code, images, videos, and general intelligence. It is positioned as a general-purpose multimodal model that can switch between dialogue, complex reasoning, and automation tasks.
Its retirement also marks that Moonshot has officially shifted its focus to Kimi K3, which was open-sourced in July. This 2.8 trillion parameter model is built using the KDA hybrid linear attention mechanism and attention residual technology, and it natively supports visual understanding with a context window of up to 1 million tokens. The official stated that K3 is the first open-source model to reach a scale of 2.8 trillion parameters, and for the past 12 months, Kimi has held the record for the largest open-source model scale for 9 of those months.