Getting started with Kimi

You can type questions in the chat box, upload files, or switch between models depending on your needs. Open kimi.com and try sending a message — say hello to Kimi:

Text

Choose a model

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

About Kimi K3

Kimi K3 is currently Kimi's most powerful model, with 2.8 trillion parameters, built on Kimi Delta Attention (KDA) and Attention Residuals, featuring native vision and a 1-million-token context window. As the world's first open 3T-class model, K3 is designed for frontier intelligence across long-horizon coding, knowledge work, and reasoning, reaching leading performance in coding, long-horizon execution, and a wide range of general intelligence tasks.

Built-in tools

Available out of the box, no configuration needed:

Visual & multimodal capabilities

Kimi K3 supports full-scenario understanding of images, videos, and office documents (PDF/Word/Excel), using a native multimodal architecture (not bolt-on OCR). It can directly parse document content and perform intermediate-to-advanced office tasks (such as writing Excel formulas or generating PPT outlines/VBA code). Combined with native vision and a 1-million-token context window, it can handle longer and more complex multimodal materials.

Multimodal input support

• Images: PNG, JPEG, WebP, GIF, etc. (via URL, Base64, or local file upload)
• Video: MP4, etc. (via video_url or image frame sequences + fps parameter)
• Documents: PDF, Word, Excel, TXT, Markdown, etc.
• Online resources: URLs, article links
• Image search: Upload an image to search for related or similar images
• Video understanding: Supports video understanding — it can analyze screen recordings and replicate interaction logic
• Visual coding: Upload screenshots or screen recordings to generate frontend code (HTML/Vue/React, etc.); can automatically reproduce web pages from design mockups

Output formats

• Plain text / mixed text-and-image: Standard conversational replies
• Markdown: Rich text with 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: Flowcharts, sequence diagrams, mind maps, etc.
• 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)

Important notes

1. File output: K3 can produce editable documents end-to-end (such as .pptx, .docx, .xlsx, .pdf) and other complex deliverables; K2.6 focuses on fast Q&A and only outputs text (such as PPT outlines, Word body text, or code).

Sessions vs. conversation turns

When you click "New Chat", think of it as starting a long conversation with Kimi.

Key differences

Session

• Chat context: Kimi refers to the content available in the current chat; long content may be compressed
• Cross-turn context: Each reply references prior conversation content as context for the next turn
• New chat: Starts a new conversation. Enabled memory can still be used for relevant questions; starting a new chat does not delete saved memories Turn
• Cumulative count: Turn 1, Turn 2, … incrementally added
• Limited context: When a conversation becomes too long, earlier details may not be fully retained, especially for long writing tasks or tasks with many requirements. This is different from memory across chats
• Context length constraint: Limited by token count; when turns accumulate, the earliest content may be compressed

Why this distinction matters

Within a single session, each new question or instruction you send is answered with reference to the available prior context. As turns accumulate and content grows longer, response quality may gradually decline.

Common scenarios

Scenario 1: Switching topics

• Wrong: Asking "help me calculate my taxes" in a session where you've been writing a 30,000-word novel
• Right: Start a new session for the tax question — avoid Kimi confusing novel characters with tax concepts Scenario 2: Long document processing
• After dozens of turns, Kimi may "forget" rules or file summaries established at the start
• Solution: Proactively summarize mid-task ("Summarize our plan so far"), or start a new session Scenario 3: Understanding product limits
• Context limits are measured in tokens, not a fixed number of turns. A long message or file can use a large amount of context in a single turn
• If you are prompted to start a new chat because the context limit has been reached, summarize key conclusions before continuing there

Best practices