Kimi K3
kimi-k3Kimi K3 by Kimi: 1M context, 128K max output, with Vision, Function Calling, Reasoning, Prompt Caching, Response Schema, Video, File Input, Streaming, Session, Web Search. Via WandAPI, Input (prompt) costs 12 Credits / M tokens, 40% below list price. Compatible with Anthropic Messages / OpenAI Responses: swap the base URL and you are connected.
Capabilities
Listed exactly as the catalog flags them, including what this model does not support.
Vision
Natively multimodal: captioning, classification, visual Q&A, chart and screenshot understanding. Images are converted to tokens at the input rate.
Function Calling
Connect the model to external tools and APIs: it decides when to call which function and with what arguments, your code runs it. Force or forbid calls with tool_choice.
Reasoning
The model thinks before answering, which markedly improves multi-step planning and reasoning for coding, math and data analysis. Thinking tokens bill at the output rate.
Prompt Caching
Repeated prefixes (system prompts, long documents) hit the cache: reads cost far less than input, writes are priced by retention time.
Response Schema
Constrain output to a JSON Schema you provide for predictable, type-safe results without parsing free text.
Video
Video input understanding or video generation.
File Input
Upload PDFs and other documents directly; the model reads the whole document with native vision.
Streaming
Tokens arrive over SSE as they are generated; the first frame is forwarded immediately with no buffering. Claude Code, Codex and similar tools depend on this.
Web Search
Vendor-hosted live web retrieval; the model decides when to search and cites sources.
Session
Server-side multi-turn context (e.g. previous_response_id in Responses) so you need not resend history.
AudioNot supported
Audio input or speech output.
RealtimeNot supported
Bidirectional low-latency voice / text sessions over WebSocket.
MediaNot supported
Image / video / music generation endpoints.
Multipart FormNot supported
multipart/form-data file-upload requests.
Binary ResponseNot supported
Endpoints that return raw audio / image bytes.
Protocols & code samples
Change the base URL to WandAPI, keep everything else. Native protocols pass straight through; “adapted” ones are translated by the gateway.
Anthropic SDK or any Messages-protocol client: point base_url at https://api.wandapi.com and use a WandAPI key.
curl https://api.wandapi.com/v1/messages \
-H "x-api-key: $WANDAPI_KEY" \
-H "anthropic-version: 2023-06-01" \
-H "content-type: application/json" \
-d '{
"model": "kimi-k3",
"max_tokens": 1024,
"messages": [{"role": "user", "content": "Explain prompt caching in one sentence"}]
}'Full price list
List price → WandAPI price per channel, every pricing dimension, nothing hidden.
| Item | Unit | Kimi list price | Kimi (Discount Channel) ×0.640% off |
|---|---|---|---|
| Input (prompt) | per 1M tokens | 12 Credits | |
| Output (completion) | per 1M tokens | 60 Credits | |
| Cache read | per 1M tokens | 1.2 Credits |
“Off” is computed at 1 Credit = ¥1 and today’s rate: 1 − channel multiplier × (1 Credit in the list currency).
FAQ
How much does the Kimi K3 API cost?
Via WandAPI, Input (prompt) costs 12 Credits / M tokens, with 1 Credit = ¥1. That is 40% below the Kimi list price.
What are the context window and output limit of Kimi K3?
Context window 1M tokens, up to 128K tokens per response.
Does Kimi K3 support function calling and structured outputs?
Function calling with tool_choice control is supported; structured outputs (JSON Schema) are supported.
How do I use Kimi K3 in Claude Code?
Set ANTHROPIC_BASE_URL=https://api.wandapi.com, ANTHROPIC_API_KEY to your WandAPI key and ANTHROPIC_MODEL=kimi-k3, then run claude. No config files to edit.
One API key for Kimi K3 and 30+ models
Free credits on sign-up, pay as you go, no monthly fee.