Moonshot AI Signals Multimodal Expansion as ‘Kiwi-do’ Model Surfaces in Testing
A mysterious artificial intelligence model labeled "Kiwi-do" has appeared on the Large Model Arena (LMArena) benchmarking platform, sparking widespread speculation that Beijing-based Moonshot AI is preparing to launch a new multimodal system in the first quarter of 2026.
According to a report originally published by Node AI View and distributed via 36Kr, market observers and technical users have identified the new entry as a potential test version of Moonshot AI’s anticipated "K2-VL" model. When queried directly by users, the Kiwi-do model reportedly identified itself as "Kimi," the consumer-facing brand of Moonshot AI.
This development aligns with recent disclosures from Moonshot AI’s leadership regarding its product roadmap. Founder Yang Zhilin recently confirmed on Reddit that a Vision-Language (VL) model was undergoing testing. Furthermore, in an internal letter released on December 31, 2025, Yang outlined plans to release new models in 2026. Citing earlier reports from STAR Market Daily, the 36Kr article notes that Kimi plans to launch a multimodal model—potentially designated K2.1 or K2.5—during the first quarter of this year.
Performance and Capabilities
Early user tests on the LMArena platform suggest Kiwi-do possesses enhanced visual and logical reasoning capabilities compared to Moonshot AI’s existing offerings.
The report highlights a comparative test involving Scalable Vector Graphics (SVG) generation. Kiwi-do successfully rendered complex requests, such as a pelican riding a bicycle and a game controller, outperforming the currently available K2-Thinking model.
Furthermore, users subjected the model to the Visual Physics Comprehension Test (VPCT), a benchmark requiring AI to apply physical laws to visual data. The report states that Kiwi-do correctly solved all presented problems, reinforcing the theory that it represents a significant upgrade in multimodal processing, which integrates text and visual inputs.
Competitive Landscape and Financial Position
The potential release comes as Moonshot AI seeks to consolidate gains made in the latter half of 2025. After facing intense competition from rivals like DeepSeek early last year, the company rebounded with the July 2025 release of Kimi-K2, a trillion-parameter open-source model utilizing a Mixture-of-Experts (MoE) architecture.
By December 2025, the LMArena open-source leaderboard ranked Moonshot AI’s "Kimi-K2-thinking" second, surpassing DeepSeek v3.2. This model introduced capabilities allowing for simultaneous "thinking and tool collaboration" and expanded context length to 256K tokens.
Financially, the company appears well-positioned to sustain its R&D spending. In his recent internal communication, Yang revealed that Moonshot AI holds cash reserves exceeding RMB 10 billion yuan (US$1.38 billion). This capital buffer distinguishes the firm from domestic competitors such as Zhipu AI and MiniMax, which are reportedly accelerating efforts to secure initial public offerings.
Yang has signaled a more measured approach to capitalization, emphasizing a distinct path toward Artificial General Intelligence (AGI) over immediate IPO preparations. The introduction of a robust multimodal agent could significantly impact the competitive landscape for AI applications in sectors ranging from intelligent customer service to healthcare and education.