Follow ZDNET: Add us as your preferred source Google.
Important points of ZDNET
- Moonshot debuted its open source Kim K2.5 model on Tuesday.
- You can generate web interfaces based solely on images or videos.
- It also includes the “agent swarm” beta feature.
Alibaba-backed Chinese AI startup Moonshot released Kim K2.5 on Tuesday, describing it in a blog post as the world’s “most powerful open source model to date.”
Building on the Kim K2 LLM that debuted last summer, Moonshot’s latest model comes with coding features that could make it a strong competitor to its own models. According to data published by Moonshot, Kimi K2.5 scored on par with Frontier models from OpenAI, Google, and Anthropic on the SWE-Bench Verified and SWE-Bench multilingual coding benchmarks.
However, the ability to create a front-end web interface from visual input is what truly differentiates it from other products.
Coding with a vision
According to Moonshot, Kim K2.5 is pre-trained with 15 trillion text and visual tokens, making it a “native multimodal model” that can generate web interfaces with interactive elements and scrolling effects from uploaded images and videos.
In a demo video of this “Coding with Vision” feature included in Moonshot’s blog post, Kimi K2.5 generated a draft of a new website based on a recorded video of an existing website, viewed from the perspective of the user’s scrolling screen. The model was able to recreate the overall aesthetic, even if it made some visual missteps along the way, such as depicting the Earth’s continents as amorphous blobs in classic AI style.
It’s unclear how practical this kind of functionality will be. (Why should companies create slightly less visually appealing AI-generated copies of already perfectly streamlined websites?) Still, generating mockups of websites and apps from images and videos alone would be a meaningful step forward for so-called “vibe coding” tools. The tool is based on an intuitive method that is easy to implement, even for non-experts, rather than traditional coding.
ChatGPT, Claude, and Gemini can generate raw code for new web assets based on screenshots and other images, but users still need to convert it into a finished, usable product. The novelty (and potential market value) of Moonshot’s new model is that it eliminates that intermediate step. “K2.5 improves image/video-to-code generation and visual debugging by inferring images and videos, lowering the barrier for users to express intent visually,” the company said in a blog post.
Also: Vibecode a Mac app in 8 hours using Claude Code, and it was more hard work than magic
If it proves useful in the real world, especially among enterprises, other developers will likely follow suit by including similar functionality in their own models.
Kimi K2.5’s coding capabilities are available through an open source platform called Kimi Code, which can be accessed through integrated development environments (IDEs) such as Cursor, VSCode, and Zed. New models are also available on Kimi.com, Kimi App, and Kim API.
swarm of agents
Moonshot also announced a research preview called “Agent Swarm,” which coordinates up to 100 “subagents” to improve performance for specific multi-step tasks.
By running multiple tasks in parallel with each other, agent swarms can also speed up computing processes. “Running these subtasks simultaneously significantly reduces end-to-end latency compared to running the agents sequentially,” Moonshot wrote in a blog post, adding that internal evaluations showed that end-to-end runtime (the total process from input to final output completion) can be reduced by up to 80%.
Also: Using Claude Code, I finished vibrating my Apple Watch app in just 12 hours instead of 2 months.
Users with an active “Allegretto” or “Vivace” Moonshot account ($31/month and $159/month, respectively) can click on the model drop-down menu at the bottom right of the prompt box,[K2.5 Agent Swarm (ベータ)]You can try Agent Swarm on the Kimi website by selecting .
