- Published on
Moonshot's First-Generation Trillion-Parameter Multimodal Model Kimi K2.5 Officially Retired at Month's End
- Authors
- Name
- DeepHorus
Moon's Dark Side recently announced that its first-generation trillion-parameter multimodal model, Kimi K2.5, will officially retire at the end of this month. The model was launched and open-sourced in January this year, and the Kimi intelligent assistant K2.5 version went live simultaneously, with the K2 model in the user interface automatically upgraded to K2.5. As Moon's Dark Side's previous smartest model, Kimi K2.5 achieved state-of-the-art performance in the open-source field across multimodal tasks such as agents, code, images, and videos.
Kimi K2.5 adopted a native multimodal architecture design, capable of processing both visual and text inputs simultaneously, supporting thinking and non-thinking modes, and adapting to various task scenarios like conversations and agents. In terms of model positioning, it was defined as a general-purpose multimodal model, capable of flexible switching between conversations, complex reasoning, and automated tasks. This design made it the most comprehensive model Moon's Dark Side had produced to date.
The retirement of Kimi K2.5 marks Moon's Dark Side's continued advancement in model iteration. In July this year, the company officially open-sourced its new-generation Kimi K3 model, which has a parameter scale of 2.8 trillion, making it the first open-source model to reach this level. K3 is built on the KDA hybrid linear attention mechanism and attention residual techniques, natively supports visual understanding, and features a 1-million-token context window.
Moon's Dark Side pointed out that over the past 12 months, its models have maintained the scale upper limit of open-source models for 9 months, demonstrating the company's leading position in large model research and development. The release of K3 further consolidates this advantage, providing more powerful foundational capabilities for multimodal AI applications. The retirement of K2.5 is expected to pave the way for the popularization and optimization of K3 and more advanced models in the future.
From a technological development perspective, the evolution from Kimi K2.5 to K3 reflects the rapid increase in parameter scale and capabilities of multimodal AI models. The parameters have grown from the trillion level to 2.8 trillion, while introducing new attention mechanisms, which are expected to bring significant improvements in complex task processing and long-text understanding. As models continue to upgrade, Moon's Dark Side's competitiveness in the AI assistant and automated services sectors will also strengthen.
For the industry, Moon's Dark Side's model iteration strategy reflects the vibrant competition in the open-source ecosystem. By continuously launching larger-scale and more advanced open-source models, the company has driven technological progress across the entire AI community. The retirement of K2.5 and the rise of K3 together outline the development trajectory of Chinese AI enterprises in the large model race.
Looking ahead, Moon's Dark Side is likely to continue advancing model innovation, exploring larger parameter scales and more efficient architecture designs. With the widespread adoption of K3, its performance in areas such as multimodal understanding and long-context processing will be tested by the market. At the same time, the retirement of K2.5 also reminds users to keep up with technology upgrades to fully utilize the latest capabilities brought by AI models.
Overall, through the retirement of Kimi K2.5 and the succession by K3, Moon's Dark Side has completed another significant technological iteration. This process not only optimizes the company's product line but also provides users and developers with more powerful tools. In the context of rapid AI technology evolution, continuous model updates will be key for enterprises to maintain competitiveness.
