Kimi K3
Moonshot AI
Moonshot AI's Kimi K3 reasoning model with a context window of about 1M tokens.
Reasoningus-eu-hosted
moonshotai/kimi-k3
Also callable as kimi-k3
Call it
Base URL https://api.thetoken.dev/v1, your API key in TOKEN_API_KEY. This model reasons before it answers; stream the response to see output as it arrives.
from openai import OpenAIimport os
client = OpenAI(base_url="https://api.thetoken.dev/v1", api_key=os.environ["TOKEN_API_KEY"])stream = client.chat.completions.create( model="moonshotai/kimi-k3", messages=[{"role": "user", "content": "Explain failover in one sentence."}], stream=True,)for chunk in stream: if chunk.choices and chunk.choices[0].delta.content: print(chunk.choices[0].delta.content, end="")import OpenAI from "openai";
const client = new OpenAI({ baseURL: "https://api.thetoken.dev/v1", apiKey: process.env.TOKEN_API_KEY });const stream = await client.chat.completions.create({ model: "moonshotai/kimi-k3", messages: [{ role: "user", content: "Explain failover in one sentence." }], stream: true,});for await (const chunk of stream) process.stdout.write(chunk.choices[0]?.delta?.content ?? "");curl https://api.thetoken.dev/v1/chat/completions \ -H "Authorization: Bearer $TOKEN_API_KEY" \ -H "Content-Type: application/json" \ -d '{"model": "moonshotai/kimi-k3", "messages": [{"role": "user", "content": "Explain failover in one sentence."}]}'import anthropicimport os
client = anthropic.Anthropic(base_url="https://api.thetoken.dev", api_key=os.environ["TOKEN_API_KEY"])message = client.messages.create( model="moonshotai/kimi-k3", max_tokens=1024, messages=[{"role": "user", "content": "Explain failover in one sentence."}],)print(message.content)from langchain_openai import ChatOpenAIimport os
llm = ChatOpenAI(model="moonshotai/kimi-k3", base_url="https://api.thetoken.dev/v1", api_key=os.environ["TOKEN_API_KEY"])print(llm.invoke("Explain failover in one sentence.").content)Providers
Requests go to the first provider and move down the list if it fails. TOKEN's price applies whichever provider serves the request.
| # | Provider | Precision (as the provider states it) | Data location | us-eu-hosted |
|---|---|---|---|---|
| 1 | DeepInfra | MXFP4 native · checked October 19, 2026 | US | Yes |
| 2 | Fireworks AI | MXFP4 native · checked October 19, 2026 | Not stated | No |
| 3 | Novita AI | Not stated (backup only) · checked October 19, 2026 | US | Yes |
Performance
Measured on TOKEN's streamed requests; shown from 10 requests in a window.
| Window | TTFT p50 | TTFT p90 | Output speed p50 | Output speed p90 |
|---|---|---|---|---|
| Last 24 hours | … | … | … | … |
| Last 7 days | … | … | … | … |
TTFT is the time from TOKEN receiving the request to the first token. Output speed is tokens per second after the first token. p50 is the median; p90 is the time 9 in 10 requests beat.
Availability
Request success: the share of real requests answered, after any failover; errors count against it. Last 30 days, one bar per day:
- 24 hours: …
- 7 days: …
- 30 days: …
Endpoint availability: a request to this model every 5 minutes from outside our infrastructure (Better Stack), on the status page. Method: Reliability.
Capabilities
| Input | text, image, video |
|---|---|
| Output | text |
| Reasoning | Yes, adjustable effort |
| Tool calling | Yes |
| Structured output (JSON schema) | Yes |
| Maximum output | 1,048,576 (default 131,072) |
| Context window | 1,040,000 tokens |
As stated by Moonshot AI, checked October 24, 2026: www.kimi.com, www.kimi.com, platform.kimi.ai, huggingface.co.
Questions
Who makes Kimi K3?
Kimi K3 is made by Moonshot AI and was released on July 16, 2026. TOKEN serves the maker's released weights through upstream providers; it does not train or modify the model.
What is the context window of Kimi K3?
1,040,000 tokens (1M). The maker states a maximum output of 1,048,576 (default 131,072) tokens.
At what precision does TOKEN serve Kimi K3?
The first provider in the routing order serves MXFP4, the maker's own native release format. TOKEN never routes to a quantized copy; a provider that does not state its precision is used only as a labelled backup.
Which providers serve Kimi K3 on TOKEN?
DeepInfra, Fireworks AI and Novita AI, in that routing order. If one fails, the request moves to the next.
How much does Kimi K3 cost?
$3.14 per 1M input tokens and $15.68 per 1M output tokens (cached input $0.32), the same whichever provider answers. You pay from prepaid credit in USD; each top-up carries a 5.5% credit purchase fee, shown at checkout.
Does Kimi K3 support tool calling, images or reasoning?
Tool calling: Yes. Image input: Yes. Reasoning: Yes, with adjustable effort. Structured (JSON schema) output: Yes.
Where are my requests processed?
At the provider that answers: US and not stated by the provider. Keys in the us-eu-hosted group only use US and EU hosts.
Does TOKEN store my prompts or responses?
No. TOKEN keeps token counts and request metadata for billing and logs, not the content of prompts or responses. See the privacy policy.
What happens if a provider of Kimi K3 fails?
TOKEN retries the request on the next provider in the routing order, up to three times, without any change on your side.
Will Kimi K3 be retired without notice?
No. TOKEN gives at least 30 days of notice before a model is retired.