Kimi K2.5

What is Kimi K2.5?

Kimi K2.5 is an open-source, multimodal trillion-parameter model built to handle huge inputs and coordinate autonomous agent swarms. It understands text, images, videos, and PDFs in one unified model, generates production-ready UI code from screenshots, and keeps context across very long documents with a 256K window. The model is designed for cost-efficient inference and can be self-hosted or run via a simple CLI, with INT4 options for local privacy-sensitive deployments. It’s aimed at developers and teams who want powerful multimodal and agentic capabilities without proprietary lock-in.

Key features

• Processes text, images, videos, and PDFs through one unified model.
• Handles extremely long inputs using a 256K-token context window.
• Coordinates autonomous agent swarms to run complex parallel workflows efficiently.
• Generates production-ready UI code directly from screenshots and design mockups.
• Open-weight, self-hostable model with INT4 options for local private deployment.
• Operates at low inference cost—about $0.39 per million input tokens.