kimi-k3
designed for frontier intelligence scenarios including long-horizon coding, knowledge work, and reasoning.
Modalities
In / Out Price
$3/ 1M
$3 Official price
$15/ 1M
$15 Official price
Context
1.048576M
Released
Jul 16, 2026
API call examples
POST /v1/chat/completionscurl --request POST 'https://api.tokatlas.com/v1/chat/completions' \
--header "Authorization: Bearer $UNIROUTER_API_KEY" \
--header 'Content-Type: application/json' \
--data '{
"model": "kimi-k3",
"messages": [
{
"role": "user",
"content": "Explain quantum entanglement."
}
],
"temperature": 0.7
}'Providers & routes
Current provider · Moonshot
Platform configured price · Group: tec_kimi
USD / 1M tokens. Each row shows input and output from the same tier. Displayed rates use the lowest available group multiplier; your API key group may differ.
Kimi K3 for large repositories and media sources
Use it for tasks that combine long reference materials with code. Ask for a source index, conclusions, and hypotheses to verify.
Long multimodal inputs
Group code, documents, and media and keep an index for each source.
Long-running development
Verify changes with test output, module lists, and acceptance criteria.
Tool calls
Keep tool results and reproducible inputs in the application.
JSON reporting
Return sources, conclusions, and missing information in fixed fields.
About
Long multimodal context
K3 uses text, images, and video to produce text. About one million tokens are shared between input and generation, with a 131,072 default output cap. Index long media and preserve source paths for repository work.
Tool-driven and structured work
Effort can be low, high, or max. Upstream tools and JSON let workflows feed back queries and tests. Validate structured fields against a schema and handle interruption and retries in the application.
API migration details
The reference distinguishes regular streaming, tool choice, and partial responses. Verify relevant fields before migration. When comparing with K3.1, keep sources, prompts, and acceptance criteria fixed.
Index a repository briefly before focusing on affected modules. When reading data early through partial or streamed responses, wait for complete, valid JSON or function arguments before using them.
kimi-k3 FAQ
Can Kimi K3 analyze a short video?
Native K3 can use video as input for a written response. Check whether the local route exposes video input and which limits apply before uploading.
How should I choose a Kimi K3 reasoning level?
Try low for straightforward extraction and evaluate high or max for cross-file analysis. Compare latency and the need for retries.
Can I keep every OpenAI SDK parameter when migrating to Kimi K3?
Keep the familiar message structure where supported, then check reasoning, tool-call and video-input fields. Validate a full turn that includes a tool result.
Can I deploy Kimi K3 myself?
Moonshot provides K3 weights. Review the license and inference requirements before hosting them, or evaluate latency and output quality through the hosted route first.