China’s Moonshot releases a new open-source model Kimi K2.5 and a coding agent

Moonshot AI, a Chinese startup supported by major players like Alibaba and HongShan (formerly Sequoia China), has just introduced its latest open-source model, Kimi K2.5. This innovative model boasts the ability to comprehend text, images, and videos, thanks to its training on a staggering 15 trillion mixed visual and text tokens, making it inherently multimodal. The capabilities of Kimi K2.5 extend into coding, where it excels at managing agent swarms—an approach that allows multiple agents to collaborate effectively. In recent benchmarks, Kimi K2.5 has demonstrated performance that not only rivals existing proprietary models but also surpasses them in specific areas. For example, in the SWE-Bench Verified benchmark, it outshines Gemini 3 Pro and achieves higher scores than both GPT 5.2 and Gemini 3 Pro on the SWE-Bench Multilingual benchmark. In terms of video comprehension, Kimi K2.5 sets a new standard by outperforming GPT 5.2 and Claude Opus 4.5 on the VideoMMMU benchmark, which evaluates a model's reasoning capabilities over videos. Moonshot AI emphasizes that while Kimi K2.5 is adept at understanding text, it can also process images or videos, enabling users to request the development of similar interfaces based on the provided media. To maximize the utility of these coding features, Moonshot AI has launched Kimi Code, an open-source coding tool designed to compete with offerings like Anthropic's Claude Code and Google's Gemini CLI. Developers can access Kimi Code via their terminals or integrate it with popular development environments like VSCode, Cursor, and Zed. This tool allows for image and video inputs, reflecting the growing trend of coding tools that are rapidly becoming major revenue generators for AI companies. Notably, Anthropic reported that Claude Code achieved an impressive $1 billion in annualized recurring revenue (ARR) last November, and by early this month, it was projected to contribute an additional $100 million by the end of 2025. Meanwhile, Moonshot's rival, Deepseek, is anticipated to launch its own advanced coding model next month, increasing competition in this burgeoning field. Founded by former Google and Meta AI researcher Yang Zhilin, Moonshot AI has raised significant funding, including $1 billion during its Series B round, establishing a valuation of $2.5 billion. Recent reports indicate that it secured an additional $500 million last month, bringing its valuation to $4.3 billion, with plans to pursue further funding at a projected $5 billion valuation.