A session is where everything begins

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

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

Choose a model

Kimi offers multiple model options. Click the model switcher above the input box to switch models.

About Kimi K3

Kimi K3 is Kimi’s most powerful model to date. With 2.8 trillion parameters, it is built on Kimi Delta Attention (KDA) and Attention Residuals, and features native vision capabilities and a 1 million-token context window. As the world’s first open 3T-class model, K3 is designed for frontier AI scenarios such as long-horizon coding, knowledge work, and reasoning. It delivers leading performance in coding, long-horizon execution, and a wide range of general intelligence tasks.

Built-in official tools

Ready to use without any additional setup:

Vision and multimodal capabilities

Kimi K3 understands images, videos, PDF/Word/Excel office documents, and more across a wide range of scenarios. Built with a native multimodal architecture (not an add-on OCR module), it can parse document content directly and perform mid- to advanced-level office tasks, such as writing Excel formulas and generating PPT outlines or VBA code. Combined with native vision capabilities and a 1 million-token context window, it can handle longer and more complex multimodal materials.

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 + the fps parameter)
• Documents: PDF, Word, Excel, TXT, Markdown, and more
• Online resources: URLs (Uniform Resource Locators) and Official Account article links
• Image-to-image search: Upload an image and call the image retrieval tool to search for related or similar images
• Video understanding: Understand videos, analyze screen recordings, and reproduce interaction logic
• Visual programming: Upload screenshots or screen recordings to generate frontend code (HTML/Vue/React, etc.), and automatically recreate webpages from design drafts

Session output formats

• Plain text / mixed text and images: standard chat replies
• Markdown: rich text, headings, lists, tables, and links
• Code blocks: HTML/Vue/React/Python/JSON/Mermaid, etc. (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 (reasoning process) and content (final answer)
• Tool call format: Function Calling standard format (for developers)

Notes

1. File generation: K3 can produce complex, end-to-end editable documents such as .pptx, .docx, .xlsx, and .pdf; K2.6 is designed more for quick Q&A and outputs text only, such as PPT outlines, Word body text, and code.

Sessions vs. turns

When you click “New session”, think of it as starting a long conversation with Kimi.

Key differences

Session

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

Turn

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

Why the distinction matters

Within the same session, every time you enter a new question or instruction, Kimi’s reply in the current turn refers to previous turns. When there are too many turns and too much accumulated content, answer quality may decline.

Common 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 you set at the beginning or the summary of an uploaded file
• What to do: proactively summarize midway through important tasks (“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 plain terms you can think of them as “turns”
• If you see a “start a new session” prompt, it means the current session has too many turns and you need to open a new one

Best practices