.png)
April 28, 2025
The Chinese startup Moonshot AI has introduced a new open-source AI model called Kimi-VL, which efficiently processes text, images, and videos. Particularly noteworthy is the model's ability to understand long documents, complex arguments, and user interfaces.
Kimi-VL is based on a Mixture-of-Experts architecture, where only a part of the model is activated for each task. With only 2.8 billion active parameters – significantly fewer than many large language models – Kimi-VL achieves results in various benchmarks that are comparable to significantly larger systems.
The model has a maximum context window of 128,000 tokens, sufficient to process an entire book or a long video transcript. According to Moonshot AI, Kimi-VL performs consistently well in tests such as LongVideoBench and MMLongBench-Doc.
The image processing capabilities of Kimi-VL are also noteworthy. Unlike some other systems, it can analyze full screenshots or complex graphics without having to break them down into smaller parts. The model also processes mathematical image problems and handwritten notes. In one test, it analyzed a handwritten manuscript, identified references to Albert Einstein, and explained their relevance.
Furthermore, the system acts as a software assistant that interprets graphical user interfaces and automates digital tasks. Moonshot AI claims that Kimi-VL outperformed many other systems, including GPT-4o, in tests where the model navigated browser menus or changed settings.
Compared to other open-source models like Qwen2.5-VL-7B and Gemma-3-12B-IT, Kimi-VL appears more efficient. According to Moonshot AI, it leads in 19 out of 24 benchmarks, despite running with significantly fewer active parameters. On MMBench-EN and AI2D, it allegedly achieves values that are normally achieved by larger, commercial models.
The company largely attributes this performance to its training approach. In addition to standard supervised fine-tuning, Kimi-VL uses reinforcement learning. A specialized version called Kimi-VL-Thinking was trained to go through longer thought steps, which increases performance on tasks that require more complex thinking, such as mathematical reasoning.
Despite the impressive performance, Kimi-VL also has limitations. The current size limits performance on very language-intensive or niche tasks, and there are continuing technical challenges with very long contexts, even with the extended context window.
Moonshot AI plans to develop larger model versions, incorporate more training data, and improve fine-tuning. The company's long-term goal is to create a "powerful yet resource-efficient system" suitable for real-world use in research and industry.
Earlier this year, Moonshot AI already released Kimi k1.5, a multimodal model for complex reasoning that, according to the company, can compete with GPT-4o in benchmarks. Kimi k1.5 is available via the Kimi.ai web interface. A demo of Kimi-VL can be found on Hugging Face.
