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

China’s Moonshot AI, which is backed by the likes of Alibaba and HongShan (formerly Sequioa China), contiguous released a caller open-source model, Kimi K2.5, which understands text, image, and video.

The institution said that the exemplary was trained connected 15 trillion mixed ocular and substance tokens, and that’s wherefore it is natively multimodal. It added that the models are bully astatine coding tasks and handling cause swarms
— an orchestration wherever aggregate agents enactment together. In released benchmarks, the exemplary matches the show of the proprietary peers and adjacent beats them successful definite tasks.

For instance, successful the coding benchmark, the Kimi K2.5 outperforms Gemini 3 Pro astatine the SWE-Bench Verified benchmark, and scores higher than GPT 5.2 and Gemini 3 Pro connected the SWE-Bench Multilingual benchmark. In video understanding, it beats GPT 5.2 and Claude Opus 4.5 connected VideoMMMU (Video Massive Multi-discipline Multimodal Understanding), a benchmark that measures however a exemplary reasons implicit videos.

Image Credits: Moonshot AI

Moonshot AI said that connected the coding front, portion the exemplary tin recognize substance well, users tin besides provender it images oregon videos and inquire it to marque a akin interface shown successful those media files.

To fto radical usage these coding capabilities, the institution has launched an open-source coding instrumentality called Kimi Code, which would rival Anthropic’s Claude Code oregon Google’s Gemini CLI. Developers tin usage Kimi Code done their terminals oregon integrate it with improvement bundle specified arsenic VSCode, Cursor, and Zed. The startup said that developers tin usage images and videos arsenic input with Kimi Code.

Coding tools person gained accelerated popularity and are becoming gross drivers for AI labs. Anthropic announced successful November that Claude Code had reached $1 cardinal successful annualized recurring gross (ARR). Earlier this month, Wired reported that by the extremity of 2025, the instrumentality had added $100 cardinal to that figure. Moonshot’s Chinese competitor, Deepseek, is acceptable to merchandise a caller exemplary with beardown coding chops adjacent month, according to a study by The Information.

Moonshot was founded by erstwhile Google and Meta AI researcher Yang Zhilin. The institution raised $1 cardinal successful backing successful a Series B circular astatine a $2.5 cardinal valuation. According to Bloomberg, the startup picked up $500 cardinal successful backing past period astatine $4.3 cardinal valuation. What’s more, the study noted that it is already seeking to rise a caller circular astatine a $5 cardinal valuation.