Kimi K2.5 — Open-Weight Multimodal Model for Agent Swarms
Moonshot AI’s Kimi K2.5 is the flagship open-weight model, continued-pretrained on ~15T multimodal tokens for long-context reasoning, visual coding, and agentic workflows.
Open-weight · 15T multimodal tokens · 256K context · Agent Swarm
What is Kimi K2.5
Kimi K2.5 is an open-weight, native multimodal model that emphasizes coding and vision, powered by a self-driven agent swarm paradigm and multiple ways to use it.
Open-Weight Continued Pretraining
Built on K2 with ~15T mixed vision+text tokens and released as open-weight.
Native Multimodal + Visual Coding
Reason over images and video for visual debugging and image/video-to-code workflows.
Agent Swarm Paradigm
Self-driven agent swarms orchestrate complex, multi-tool tasks in parallel.
Multiple Entry Points + Four Modes
Use via Kimi.com, Kimi App, API, or Kimi Code with Instant, Thinking, Agent, and Agent Swarm (Beta) modes.
Why Choose Kimi K2.5
Choose Kimi K2.5 for frontier-level agent benchmarks at lower cost, plus multimodal coding and long-context strengths.
How to Get Started with Kimi K2.5
Use Kimi K2.5 in four steps to go from install to production-ready multimodal agents.
Install Kimi CLI
Run the official install script, verify with `kimi --version`, then launch `kimi` in your project.
Choose Your Deployment Mode
Deploy Kimi K2.5 via Kimi.com or Kimi App for instant use, or connect via API / Kimi Code for engineering workflows. Web & app include Instant, Thinking, Agent, and Agent Swarm (Beta).
Explore Multimodal Capabilities
Use Kimi K2.5 image/video reasoning for visual debugging and image/video-to-code tasks, plus long-context analysis.
Scale Your AI Applications
Leverage 256K context and Agent Swarm parallelism (up to 100 sub-agents, 1,500 tool calls, up to 4.5× speedup).
Key Features of Kimi K2.5
Kimi K2.5 core capabilities for multimodal reasoning and agentic execution.
Native Multimodal Reasoning
Kimi K2.5 reasons over images and video for visual debugging and image/video-to-code tasks.
Agent Swarm Parallelism
PARL-trained swarm coordination with up to 100 sub-agents and 1,500 tool steps.
Four Operating Modes
Instant, Thinking, Agent, and Agent Swarm (Beta) for different task depths.
Kimi Code Toolchain
Terminal + IDE (VSCode/Cursor/Zed), open-source, image/video input, and MCP-friendly.
Office-Grade Productivity
Generate dense outputs: documents, spreadsheets, PDFs, and slides.
256K Context Window
Built for long-context reasoning at scale.
Developers Love Kimi K2.5
Scale and speed highlights from the official Kimi K2.5 release.
15T Multimodal pretraining tokens
Multimodal pretraining tokens
256K Context length (official eval setting)
Context length (official eval setting)
100 Sub-agents in a swarm
Sub-agents in a swarm
1,500 Parallel tool calls
Parallel tool calls
4.5× Max speedup vs single agent
Max speedup vs single agent
What Developers Say About Kimi K2.5
Common themes from Kimi K2.5 developer feedback on real-world usage.
Terminal + IDE workflows and image/video input in Kimi Code make real project adoption smoother.
Liam Zhao, Engineering Lead, Workflow Tools
Liam Zhao
Engineering Lead, Workflow Tools
OpenAI/Anthropic-compatible APIs reduce migration cost and let teams reuse existing SDKs.
Maya Chen, Platform Architect, AI Infra
Maya Chen
Platform Architect, AI Infra
Agent Swarm parallelism and long context help cut end-to-end time on complex tasks.
Victor Sun, Agent Systems Engineer
Victor Sun
Agent Systems Engineer
Native multimodal reasoning improves visual debugging and UI-to-code workflows.
Nina Park, Product Engineer, Visual AI
Nina Park
Product Engineer, Visual AI
The four modes make it easy to move from quick answers to deep agent workflows without changing tools.
Alex Rivera, Staff Engineer, Dev Experience
Alex Rivera
Staff Engineer, Dev Experience
Agent Swarm parallel calls and 256K context cut our multi-doc workflows from hours to minutes.
Grace Liu, Automation Lead, Enterprise Apps
Grace Liu
Automation Lead, Enterprise Apps
Frequently Asked Questions About Kimi K2.5
Questions about Kimi K2.5? Email us at [email protected].
Build with Kimi K2.5
From multimodal reasoning to agent swarms, start shipping with Kimi K2.5 today.
