Every session starts with a conversation

You can type questions in the chat box, upload files, or switch models as needed.

Open kimi.com, send a message, and say hello to Kimi:

Powered by K2.6 by default

The Kimi K2.6 open-source model is a comprehensive upgrade over Kimi K2.5, with a focus on long-horizon coding, Agent Swarm, and proactive agent capabilities. It reaches state-of-the-art performance among open-source models in coding, long-horizon execution, and a wide range of general intelligence tasks. Kimi K2.6 is also Kimi’s most capable model to date. Designed with a native multimodal architecture, it supports both visual and text input, Thinking and non-thinking modes, and both chat and Agent tasks.

This means Kimi is not only good at Q&A and conversation. It can also plan tasks independently, call tools, and complete multi-step reasoning, supporting deep thinking and execution for complex tasks.

Official tools built into K2.6

No additional setup is required:

Vision and multimodal capabilities

K2.6 supports full-scenario understanding of images, videos, PDF/Word/Excel office documents, and more. With a native multimodal architecture rather than an external OCR add-on, it can directly parse document content and perform intermediate to advanced office tasks, such as writing Excel formulas or generating PPT outlines/VBA code.

Supported multimodal inputs

• Images: PNG, JPEG, WebP, GIF, and more (supports URL, Base64, and local files)
• Videos: MP4 and more (supports video_url or image sequence frames + fps parameter)
• Documents: PDF, Word, Excel, TXT, Markdown, and more
• Online resources: URL (Uniform Resource Locator), WeChat official account article links
• Image-based image search: Upload an image and call the image retrieval tool to search for related or similar images
• Video understanding: Supports video understanding, can analyze screen recordings and reproduce interaction logic
• Visual programming: Supports uploading screenshots or screen recordings to generate frontend code (HTML/Vue/React, etc.), and can automatically recreate web pages from design drafts

Session output formats

• Plain text / mixed text and images: standard chat replies
• Markdown: rich text, headings, lists, tables, links
• Code blocks: HTML/Vue/React/Python/JSON/Mermaid, and more (with language identifiers)
• JSON structured data: forced JSON mode output
• Mermaid diagrams: syntax for flowcharts, sequence diagrams, mind maps, and more
• LaTeX math formulas: inline and block-level mathematical expressions
• Dual-mode output (Thinking): returns both reasoning_content (thinking process) and content (final answer)
• Tool call format: Function Calling standard format for developers

Notes

1. Image generation limitation: Standard chat mode only supports searching for images and does not support directly generating images. To generate images, use Agent mode.
2. File output limitation: Only text output is supported, such as PPT outlines, Word body text, or code. Direct binary file output, such as .pptx, .docx, .xlsx, or .pdf, is not supported.
3. To generate images or create complex editable documents, use Agent mode or combine it with a code execution tool.

Session vs. chat turn

When you click “New chat”, imagine you’re starting a long conversation with Kimi.

Key differences

Session

• Continuous memory: Kimi remembers all context within the current chat
• Cross-turn context: Each reply refers to the previous conversation as context for the next turn
• A clear reset point: Clicking “New chat” starts over and clears the historical context

Turn

• Cumulative count: Turn 1, Turn 2, and so on
• Limited memory: If there are too many turns, such as more than 50, earlier content may be “forgotten”, especially in scenarios like long-form fiction writing or tasks with clear constraints and requirements
• Limited by context length: Due to token (word count) limits, the earliest content may be compressed when there are too many turns

Why this distinction matters

Within the same session, every new question or instruction you enter becomes part of the context Kimi refers to in the current turn. When a conversation has too many turns and the accumulated content becomes too long, response quality may decline.

Typical scenarios

Scenario 1: Starting a new topic

Scenario 2: Processing long documents

• After dozens of turns in one session, the AI may “forget” the rules set at the beginning or the summary of files you uploaded
• Suggested approach: actively summarize during important tasks, for example, “Summarize our current plan,” or start a new session

Scenario 3: Understanding product limits

• Context length limits are usually calculated by token (word count), but in everyday terms you can think of them as “turns”
• If you see a prompt to “start a new session”, it means the current session has too many turns and you need to start a new one

Best practices

1. Create a new session for each task: For each independent task, such as writing a paper, researching information, or debugging code, create a separate session to avoid context interference
2. Summarize long sessions regularly: After more than 20 turns, ask the AI to summarize key information to prevent information loss
3. Clear sensitive information promptly: Chat history is retained. After finishing content involving privacy, you can delete the entire session