Moonshot AI's Kimi K2.5 to Retire End of August; 2.8 Trillion-Parameter K3 Takes Over

Moonshot AI's large language model Kimi K2.5 is nearing the end of its service life. On August 24, the official Kimi Weibo account confirmed that the first-generation trillion-parameter multimodal model will officially go offline at the end of this month. Moonshot AI had previously announced that the model's API interface would cease service on August 31.

From its January release to its end-of-August retirement, Kimi K2.5's service cycle spans only about eight months. As Moonshot AI's flagship model at the time of launch, K2.5 featured a native multimodal architecture supporting text, image, and video inputs, along with the ability to toggle between thinking and non-thinking modes and coordinate multiple sub-agents to execute complex tasks. The model achieved open-source SOTA (State of the Art) performance across dimensions including Agent capabilities, coding, image, video, and general intelligence tasks.

The retirement of K2.5 is not an isolated event but part of Moonshot AI's model iteration cadence. In July this year, the company officially open-sourced its next-generation Kimi K3 model. K3 boasts 2.8 trillion parameters, built on the KDA hybrid linear attention mechanism and attention residual technology, with native visual understanding support and a context window expanded to 1 million tokens. According to official documentation, K3 is the first open-source model to reach the 2.8 trillion parameter scale, and Kimi-series models have held the record for the largest open-source model scale in 9 of the past 12 months.

Moonshot AI's official website now showcases Kimi K3 as its new flagship model, and K2.5's retirement marks the completion of a generational shift in the product line.

Cloud service providers are also driving migration. Tencent Cloud previously announced it would officially take the Kimi K2.5 model offline at 00:00 Beijing time on August 31 and recommended users migrate to Kimi K2.6. Judging by the version numbering, K2.6 appears to be a transitional iteration within the K2 series, while K3 represents a comprehensive architectural upgrade.

From K2.5 to K3, parameter scale has jumped from the trillion level to 2.8 trillion, the context window has expanded to 1 million tokens, and the architecture has shifted from traditional attention mechanisms to KDA hybrid linear attention. This evolution reflects the core competitive logic of the current large model industry: competing for developer ecosystems and market share through faster iteration cycles, larger parameter scales, and longer context windows.

For enterprise users and developers still using the K2.5 API, the August 31 deadline means they have less than a week to complete model migration and testing. Both Moonshot AI and Tencent Cloud recommend users transition to newer versions to avoid service disruptions impacting business operations.