Back to all articles
•Technology

China's Moonshot AI Unveils Multimodal Model Kimi K2.5 to Challenge Global Leaders

View original source

China's Moonshot AI, supported by Alibaba and HongShan (formerly Sequoia China), has launched Kimi K2.5, a new open-source AI model proficient in understanding text, images, and video. The model:

  • Was trained on 15 trillion mixed visual and text tokens, making it natively multimodal.
  • Performs exceptionally in coding tasks and orchestrating agent swarms.
  • Outperforms top models in benchmarks, such as SWE-Bench Verified and VideoMMMU.

In coding capabilities, Kimi K2.5 surpasses Gemini 3 Pro and shows superior performance against GPT 5.2 and Claude Opus 4.5 in video understanding tasks. Additionally:

  • Moonshot AI has introduced Kimi Code, an open-source tool competing with Anthropic's Claude Code and Google's Gemini CLI.
  • Kimi Code is compatible with development environments like VSCode, Cursor, and Zed, and allows the use of images and videos as input.

The release comes as coding tools increasingly dominate as revenue generators for AI labs, highlighted by Anthropic's Claude Code reaching $1 billion ARR. Moonshot AI, founded by ex-Google and Meta researcher Yang Zhilin, has seen substantial investment, achieving a $4.3 billion valuation after a recent $500 million funding injection.