Recently, with Moonshot AI completing a C-round financing of 3.5 billion RMB, a mysterious new model called "Kiwi-do" has made a dazzling debut on the large model arena LmArena. Since its exposure, this new model has attracted widespread attention and discussion.

A blogger accidentally discovered "Kiwi-do" on LmArena and inquired about its source, to which Kiwi-do claimed to come from "Moonshot AI." Interestingly, the training data of this model is up to January 2025, and its performance is impressive, especially in the Visual Physics Comprehension Test (VPCT), where it demonstrated extraordinary capabilities. It has even been speculated that it might be the multimodal model K2-VL that Kimi is currently developing.

image.png

Moonshot AI previously publicly stated that they are accelerating their model development in the direction of visual and language integration. According to a report by the Science and Technology Daily, Kimi plans to launch a new multimodal model in the first quarter of this year, possibly named K2.1 or K2.5. Obviously, the emergence of Kiwi-do has made people more eager for this release.

image.png

From the test results, Kiwi-do shows significant differences from the existing K2-Thinking model in the SVG drawing test, ruling out the possibility that they are the same model. Additionally, Kiwi-do successfully passed the challenging VPCT test, demonstrating its outstanding multimodal capabilities. This capability combines visual understanding with physical reasoning, and is expected to provide strong support for commercial application scenarios such as document parsing and dashboard analysis.

With the continuous advancement of AI technology, the emergence of Kiwi-do undoubtedly opens new doors for future multimodal applications, and everyone is looking forward to its performance in practical applications.