Moonshot Vision & agentic coding
Kimi K3
Keep a long project context intact while the work moves forward.
Kimi K3 is a Moonshot flagship route with native vision and a catalog context value of up to 1M tokens. It accepts text, images and video, and is positioned for long-horizon coding, knowledge work and terminal-oriented agent tasks. Use it when a project needs multimodal context plus a deliberate reasoning and tool loop.
- Text + image + video
- 1M context
- Agentic coding
Kimi K3 is a text-output route. It can help plan and call tools when your application exposes them, but it does not include a terminal, repository or hosted agent harness by itself.
AI Chat
Send a prompt and see the response here.
kimi-k3Start a conversation with this model and ask follow-up questions.
USD /1M tokens
Pricing
| Group | Input | Output | Cache Read |
|---|---|---|---|
| default | $3 | $15 | $0.3 |
| kimi-0.9 | $2.7 | $13.5 | $0.27 |
Public list prices are shown in USD. Final charges may vary by account group and usage tier.
Overview
What is Kimi K3?
Preserve the full task history when the reasoning chain matters, but keep checkpoints and output contracts so a long conversation remains inspectable. A new session can be safer when the agent needs a clean task boundary.
Video input can provide context for review, but the result should still cite timestamps or frames. Do not treat a multimodal input flag as proof that every duration or encoding is accepted.
Keep the project graph visible
Provide the repository map, constraints and tests. Ask for a focused change and a checkpoint before any broad rewrite.
Anchor claims to frames
Ask for timestamps, regions or visible evidence when reviewing a screen recording or visual artifact.
Start a clean run
Use a new session for a bounded task, then retain tool traces and artifacts for review.
From brief to deliverable
Put Kimi K3 to work
Engineering teams
Plan a multi-file engineering change
Trace a request through the repository and produce a staged implementation plan with tests and rollback points.
- Tools you need
- Repository search and a test runner for actual verification.
- Keep in mind
- Long context helps coordination but does not replace compiling or testing the patch.
View a task brief
Use the supplied issue text, repository tree, affected files and project checks. to plan plan a multi-file engineering change. Return a reviewable change plan and acceptance matrix., state the evidence limits and list the checks required before accepting the result. Do not claim an output or interaction was verified without direct evidence.
Product and QA teams
Review a recorded workflow
Use selected frames and a transcript to identify UI or process problems, preserving timestamps for each finding.
- Tools you need
- Media extraction, a player and a browser for interaction checks.
- Keep in mind
- The model cannot establish hidden state, focus order or network behavior from a recording alone.
View a task brief
Use the supplied a short recording, transcript, expected flow and severity rules. to plan review a recorded workflow. Return a timestamped issue report with reproduction steps., state the evidence limits and list the checks required before accepting the result. Do not claim an output or interaction was verified without direct evidence.
Agent and developer-tool teams
Prepare a terminal-tool run
Return typed tool arguments, dependencies and a stop state before the application allows a terminal action.
- Tools you need
- A tool gateway, permissions service and command log.
- Keep in mind
- The model must not claim commands ran or files changed without returned tool evidence.
View a task brief
Use the supplied a task, tool schemas, permissions and allowed directories. to plan prepare a terminal-tool run. Return a validated execution plan and audit record., state the evidence limits and list the checks required before accepting the result. Do not claim an output or interaction was verified without direct evidence.
Capabilities in context
Key features
| Feature | What you get | Why it matters |
|---|---|---|
| Context capacity | 1,048,576 tokens shown in the catalog | Use long context for project continuity while keeping retrieval and checkpoints explicit. |
| Vision inputs | Text, image and video inputs are listed | Use visual evidence with timestamps or regions and verify media limits for the endpoint. |
| Reasoning mode | The catalog notes max reasoning effort for this route | Monitor latency and cost, and preserve enough history for the task without carrying unrelated work. |
| Tool collaboration | Tool use is listed by the catalog | Expose typed functions and keep authorization, execution and results in your application. |
| Model identity | TokenHot model ID: kimi-k3 | Record the route and session boundary so long runs can be reproduced. |
The page reflects TokenHot catalog modalities and route metadata. Confirm media constraints, reasoning controls and tool parameters on the endpoint you use.
Audience & fit
Where Kimi K3 fits
Engineering teams with large project context
Use Kimi K3 for code planning, repository Q&A and multi-file changes that benefit from a coherent project view.
Product and QA teams
Bring a recording and expected flow together, then reproduce findings in a browser before fixing them.
Agent-tool builders
Keep sessions bounded and tool contracts explicit. Preserve traces for every consequential action.
Cost context
Plan the cost of a useful result
Budget reasoning history
Preserving full history can improve continuity while increasing input cost. Use checkpoints and trim irrelevant turns deliberately.
Price media preparation
Video extraction, frame selection and transcription are part of the workflow cost even before the model request.
Separate tool charges
Terminal calls, storage and validation belong in the application budget. Track them separately from model tokens.
Use the current pricing table for available rates and account groups. Model IDs and prices come from TokenHot’s catalog; provider pricing and subscriptions are separate.
FAQ
Frequently asked questions
What is Kimi K3 best suited for?
Kimi K3 is suited to the input, output, and capability types listed on this page. Test production workloads before choosing it for a critical system.
How is Kimi K3 priced?
Pricing is shown in the pricing section above. Actual cost depends on usage volume, account group, and the request payload.
How can I call Kimi K3?
Use the model ID shown above with one of the supported API protocols. Code examples are provided when usage data is available.
Can Kimi K3 return video?
The catalog lists video input and text output. Use a video-generation route when the deliverable is a video.
Should every task preserve full reasoning history?
Keep the context needed for continuity, but start a new bounded session when unrelated history increases cost or risk.
Does Kimi K3 include terminal access?
No. Your application must provide tools, permissions and the execution environment, then return actual results.
What should a visual coding prompt include?
Name the frame or region to inspect, the expected behavior, the relevant code and the check that would confirm the diagnosis.
Get started
Build with Kimi K3
API