Kimi K3: specs, pricing and how to run it

Kimi K3 at a glance

Moonshot AI describes Kimi K3 as the world's first open-source model in the 3-trillion-parameter class, built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes) within a Stable LatentMoE framework. Moonshot's API docs now list K3 as its previous-generation flagship behind Kimi K3.1 Pro, and the table below covers its core architecture and API specs.

Kimi K3 API pricing

On the official Kimi API, you call the model as kimi-k3 against the https://api.moonshot.ai/v1 base URL using any standard OpenAI SDK. Kimi K3 pricing is a flat rate across the entire context window with no tiering by context length: $3.00 per million input tokens, $0.30 per million cached input tokens and $15.00 per million output tokens, plus $3.00 per million for 5-minute cache writes or $6.00 per million for 1-hour cache writes. Context caching turns on automatically with a 5-minute default TTL, and kimi-k3 unlocks on your account once you top up at least $1, with your cumulative top-up setting your rate limits.

For example, a single uncached request with 500,000 input tokens and 20,000 output tokens on kimi-k3 runs $1.50 for input plus $0.30 for output, coming to $1.80 in total. Running that exact same request on Moonshot's newer kimi-k3.1-pro model, which is priced at $2.00 per million input tokens, $0.20 cached and $10.00 output, drops the cost to $1.00 plus $0.20, or $1.20 total.

Request settings K3 fixes

• K3 always runs with thinking enabled, and reasoning_effort accepts low, high or max, with max set as the default on the Kimi API.
• Output length via max_completion_tokens defaults to 131072 and can be raised all the way up to the full 1048576 context limit.
• Sampling parameters are locked to temperature at 1.0, top_p at 0.95, n at 1, presence_penalty at 0 and frequency_penalty at 0, so leave them out of your request payload.
• During multi-turn conversations and tool calls, you must send back the complete assistant message object from the previous turn rather than just its content string.

Open weights and the Kimi K3 License

If you're looking at Kimi K3 open weights for self-hosting or custom builds, the full weights and source code are hosted on Hugging Face under the Kimi K3 License. That license lets you freely use, copy, modify, merge, fine-tune, publish, distribute, sublicense and sell the model as long as you keep the original copyright notice intact. The weights ship in MXFP4 format alongside MXFP8 activations produced through quantization-aware training, and include the 401M-parameter MoonViT-V2 vision encoder.

• Any Model-as-a-Service business whose revenue (including affiliates) passes $20 million over any 12 consecutive months has to sign a separate commercial agreement with Moonshot AI before using the model commercially.
• Any commercial product with over 100 million monthly active users or over $20 million in monthly revenue must display "Kimi K3" prominently in its user interface.
• Neither of those two restrictions applies to internal company use, or to traffic you run through Moonshot AI's official products or certified inference partners.

Kimi K3 in Kimi Code, and K3.1 Pro

• Kimi Code includes the k3 model starting on the Plus plan at $19 a month, while unlocking the full 1M context window requires the Pro plan at $39 a month or higher (see Kimi Code plans for the full breakdown).
• Kimi Code also exposes a k3-256k model variant capped at a 262,144-token context window, and running k3 with the 1M window burns through roughly twice as much plan quota as k3-256k.
• Inside Kimi Code, K3 defaults its reasoning effort to high rather than the max default used on the raw Kimi API.
• Moonshot rolled out Kimi K3.1 Pro (kimi-k3.1-pro) to Kimi Code on October 23, 2026 as its current flagship, matching the 1M-token context window while cutting direct API prices to $2.00 per million input tokens and $10.00 per million output tokens.

Call Kimi K3 through Plan2API

Plan2API lets you connect an existing Kimi membership by pasting a Kimi Code API key and calling K3 through an OpenAI-compatible POST /v1/chat/completions endpoint at https://api.plan2api.com/v1 (details on our Kimi API integration). You request the model as kimi/k3 (or its alias moonshotai/kimi-k3) with a p2a_ key in the Authorization: Bearer header, which gives you the 1,048,576-token context window: conversations start on the 256K window and automatically switch over to the 1M window once they pass about 192K tokens. The endpoint supports streaming, tool calling, JSON mode (json_object or json_schema), text and image input, and reasoning_effort set to low, high or max (defaulting to high), plus Plan2API automatically fetches public image URLs in your messages for you (up to 32 images per request, up to 20 MB and 15 seconds each). The Node example follows below.

Every request you make through kimi/k3 draws directly on your connected Kimi Code subscription quota with zero per-token billing from Plan2API, and kimi/k3 is currently the only model in Plan2API's Kimi catalog. You can run it on the Free plan at $0 with no card required for 1 connected subscription and up to 50 chat requests a day, or move to the Pro plan at $12 a month ($99 a year) for 3 connected subscriptions, 16 concurrent requests and no daily request cap.

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://api.plan2api.com/v1",
  apiKey: process.env.PLAN2API_KEY,
});

const response = await client.chat.completions.create({
  model: "kimi/k3",
  reasoning_effort: "max",
  messages: [{ role: "user", content: "Refactor this module and list the breaking changes." }],
});

console.log(response.choices[0].message.content);