At a Glance
- Moonshot AI launched Kimi K2.5, a multimodal model trained on 15 trillion tokens.
- The model tops competitor benchmarks in coding and video understanding.
- A new open-source coding tool, Kimi Code, is now available for developers.
Why it matters: The release signals a major push from China’s AI sector toward open-source, multimodal systems that rival Western offerings.
Moonshot AI announced the launch of its newest open-source model, Kimi K2.5, on January 27, 2026. The company said the model was trained on 15 trillion mixed visual and text tokens, giving it native multimodal capabilities. It also introduced Kimi Code, a coding tool that lets developers feed images or videos into code generation.
New Multimodal Model Kimi K2.5
Moonshot AI described Kimi K2.5 as a system that can understand text, image, and video simultaneously. The model is also designed to handle agent swarms, where multiple agents coordinate to complete tasks. In released benchmarks, the model matches or surpasses proprietary peers.
- Training data: 15 trillion mixed visual and text tokens
- Multimodal: text, image, video
- Special features: coding assistance, agent swarm orchestration
A short video from founder Zhilin Yang shows the model in action:
“It’s his first time speaking on camera like this, and he really wanted to share Kimi K2.5 with you!” said Kimi.ai (@Kimi_Moonshot).
Benchmark Performance
Kimi K2.5 outperforms several industry leaders in key areas.
| Benchmark | Kimi K2.5 | Competitor | Result |
|---|---|---|---|
| SWE-Bench Verified | Top | Gemini 3 Pro | Beats |
| SWE-Bench Multilingual | Higher | GPT 5.2, Gemini 3 Pro | Beats |
| VideoMMMU | Top | GPT 5.2, Claude Opus 4.5 | Beats |
The model’s coding prowess is highlighted by its performance on the SWE-Bench Verified and Multilingual tests, while its video understanding is measured by the VideoMMMU benchmark.
Coding Capabilities and Kimi Code Tool
Moonshot AI’s coding tool, Kimi Code, is designed to compete with Anthropic’s Claude Code and Google’s Gemini CLI. Developers can use it directly from terminals or integrate it with popular editors such as VSCode, Cursor, and Zed. The tool also accepts images and videos as input, enabling visual context for code generation.
- Terminal access: CLI interface
- Editor integrations: VSCode, Cursor, Zed
- Visual input: Images and videos
Coding tools are becoming significant revenue drivers for AI labs. Anthropic announced in November that Claude Code had reached $1 billion in annualized recurring revenue (ARR). Wired reported that by the end of 2025, the tool had added $100 million to that figure.
Market Context and Funding
Moonshot AI was founded by former Google and Meta AI researcher Zhilin Yang. The company raised $1 billion in a Series B round at a $2.5 billion valuation. Bloomberg reported that the startup secured $500 million last month at a $4.3 billion valuation and is now seeking a new round at a $5 billion valuation.
| Funding Round | Amount | Valuation |
|---|---|---|
| Series B | $1 billion | $2.5 billion |
| Recent round | $500 million | $4.3 billion |
| Upcoming round | – | $5 billion |
The rapid funding growth reflects investor confidence in China’s AI ecosystem.
Competitive Landscape
Moonshot’s Chinese competitor, Deepseek, is set to release a new model with strong coding capabilities next month, according to a report by The Information. This development adds pressure on Moonshot to maintain its competitive edge.
- Moonshot AI: Kimi K2.5, Kimi Code
- Deepseek: Upcoming model focused on coding
- Anthropic: Claude Code, $1 billion ARR
- Google: Gemini CLI
Key Takeaways
Moonshot AI’s Kimi K2.5 demonstrates that Chinese AI labs can produce multimodal models that rival Western offerings. With strong benchmark results and an open-source coding tool, the company is positioning itself as a major player in the global AI landscape. Continued investment and upcoming competitor releases will shape the next phase of the AI arms race.
—
