Moonshot Unveils New Open-Source Model Kimi K2.5 and Coding Agent

Moonshot AI, backed by investors including Alibaba and Sequoia China (formerly Sequoia Capital China), today unveiled its new open-source model, Kimi K2.5, which possesses capabilities to understand text, images, and video.

The company stated that this model was trained on 15 trillion mixed visual and text tokens, granting it multimodal characteristics. It further added that these models excel in coding tasks and handling agentic groups—a form of orchestration for multi-agent collaboration. In released benchmarks, the model's performance rivals, and in some tasks surpasses, that of proprietary models.

For instance, in coding benchmarks, Kimi K2.5 outperformed Gemini 3 Pro on the SWE-Bench validation benchmark and scored higher than GPT 5.2 and Gemini 3 Pro on the SWE-Bench multilingual benchmark. In video understanding, it surpassed GPT 5.2 and Claude Opus 4.5 on the VideoMMMU (Video Massive Multidisciplinary Multimodal Understanding) benchmark, which assesses a model's reasoning capabilities on video content.

Moonshot AI noted that for coding, while the model understands text effectively, users can also input images or videos and request it to generate code resembling interfaces displayed in those media files.

To facilitate the use of these coding features, the company launched an open-source coding tool named Kimi Code, positioning it to compete with Anthropic's Claude Code or Google's Gemini CLI. Developers can utilize Kimi Code via their terminal or integrate it with development software like VSCode, Cursor, and Zed. The startup stated that developers can use images and videos as input for Kimi Code.

Coding tools have rapidly gained popularity, becoming significant revenue drivers for AI labs. Anthropic announced in November that Claude Code had reached an annual recurring revenue (ARR) of $1 billion. Earlier this month, Wired reported that the tool had added another $100 million to that figure by the end of 2025. According to The Information, Moonshot's Chinese competitor Deepseek plans to release a new model with robust coding capabilities next month.

Moonshot was founded by Yang Zhilin, a former AI researcher at Google and Meta. The company raised $1 billion in a Series B funding round, valuing it at $2.5 billion. Per Bloomberg reports, the startup secured $500 million in funding last month, achieving a valuation of $4.3 billion. Furthermore, reports indicate it is already seeking a new funding round at a valuation of $5 billion.