Kimi K2.5: New Open-Source Model with Visual Agentic Intelligence & Coding

Hey everyone! Get ready to be excited because Kimi has just dropped a game-changer: Kimi K2.5. This isn't just another update; it's being hailed as their most powerful open-source model yet! Building on the foundation of Kimi K2, this new iteration has been extensively pret-trained over approximately 15 trillion mixed visual and text tokens.

K2.5 is a truly native multimodal model, meaning it's designed from the ground up to handle both text and visual information seamlessly. It's bringing state-of-the-art coding and vision capabilities, powered by a super cool self-directed agent swarm paradigm. Let's dive into what makes this model so special!

What it's for

Kimi K2.5 is engineered to tackle complex tasks with incredible efficiency and intelligence. Its core strengths lie in its visual agentic intelligence and advanced coding.

For those really tricky problems, Kimi K2.5 can orchestrate a self-directed agent swarm of up to 100 sub-agents. Imagine that many tiny AI assistants working in parallel! This swarm can execute workflows across up to 1,500 tool calls, which means it can reduce execution time by an astonishing 4.5x compared to a single-agent setup. No more waiting around!

Coding enthusiasts, listen up! Kimi K2.5 is currently the strongest open-source model for coding. It truly shines in front-end development, capable of transforming simple conversations into complete front-end interfaces, packed with interactive layouts and rich animations. And it goes beyond text — K2.5 excels at coding with vision, reasoning over images and video to significantly improve image/video-to-code generation and visual debugging. This means expressing your design intent visually just got a whole lot easier.

Why it matters

This model isn't just about cool features; it's about shifting how we interact with AI and tackle demanding projects. The self-directed agent swarm paradigm means unparalleled efficiency for complex tasks, automating what used to take ages. Imagine developers creating stunning, interactive UIs from a few prompts, or quickly debugging visual components just by showing the model an image or video!

Kimi K2.5's robust performance, especially across agentic benchmarks like HLE, BrowseComp, and SWE-Verified, means you're getting top-tier capabilities. What's more, it delivers this power at a fraction of the cost, making advanced AI tools more accessible for everyone.

Where you get it

Ready to give Kimi K2.5 a spin? You can access this powerhouse through several convenient channels: Kimi.com, the Kimi App, the API, and Kimi Code.

On Kimi.com and the Kimi App, you'll find four distinct modes to suit your needs: K2.5 Instant, K2.5 Thinking, K2.5 Agent, and the exciting K2.5 Agent Swarm (Beta). Currently, the Agent Swarm (Beta) is available on Kimi.com, and high-tier paid users can even snag some free credits to experience it!

Read more: Kimi K2.5 Announcement Blog for details and how to get started on the Kimi Official Website.