Moonshot AI drops Kimi K2.5, topping Gemini 3 Pro and GPT-5.2
Moonshot AI just launched Kimi K2.5, its latest open-source multimodal model, trained on a massive 15 trillion visual and text tokens.
Early tests say it outperforms both Gemini 3 Pro and GPT-5.2 in key benchmarks, especially for multilingual tasks.
Why Kimi K2.5 stands out
Kimi K2.5 can turn images or videos straight into interactive web UIs using its open-source Kimi Code—with Kimi Code available for use in IDEs—and it works with tools like VSCode, Cursor, and Zed.
Its new Agent Swarm feature lets up to 100 sub-agents handle tasks at once, cutting processing time by up to 80% in Moonshot's internal evaluations.
How it compares (and what it costs)
Unlike ChatGPT or Claude, which need extra steps to build UIs from visuals, Kimi Code cuts out that intermediary step—putting it head-to-head with Anthropic's Claude Code and Google's Gemini CLI.
You can use the model on platforms like Kimi.com, Hugging Face; but if you want Agent Swarm powers, you'll need a paid plan starting at $31/month.