Kimi K3

Kimi K3 is Moonshot AI's open-weight flagship multimodal model for long-horizon coding, knowledge work, and reasoning, with a 1M-token context window and official API access.

Platform: Replicate

Language ModelMultimodalCoding AssistantAgent WorkflowsLong ContextMoonshot AI

Kimi API / Kimi Code k3 or k3-256k / Kimi Work / self-hosted 64+ accelerator supernode via vLLM, SGLang, or TokenSpeed

Kimi international API: $0.30 per million cache-hit input tokens, $3 per million cache-miss input tokens, and $15 per million output tokens; self-hosting is governed by the Kimi K3 License

Commercial

🚀Function Overview

Kimi K3 is a 2.8-trillion-parameter sparse Mixture-of-Experts model with native vision, a 1-million-token context window, official API access, and a full open-weight checkpoint under the Kimi K3 License.

Key Features

Official Kimi API model ID kimi-k3
Kimi Code model IDs k3 and k3-256k, with k3-256k fixed to a 256K context window
1-million-token context window
Native vision for text, image, and video-oriented workflows
2.8 trillion total parameters with 16 of 896 experts activated
104 billion activated parameters per token
Kimi Delta Attention and Attention Residuals architecture
MXFP4 checkpoint weights with MXFP8 activations from quantization-aware training
Low, high, and max reasoning effort on the Kimi API, with max as the default
Kimi Code documents high as the default reasoning effort for both k3 and k3-256k
Official international API pricing for cache-hit input, cache-miss input, and output tokens
Official 1.56 TB checkpoint with 96 safetensor shards under the Kimi K3 License

Use Cases

•Long-horizon software engineering through Kimi Code or an API agent harness
•Repository and document analysis using a 1M-token context window
•Multimodal reasoning over text, images, screenshots, and video
•Knowledge-work agents for research, spreadsheets, presentations, and reports
•Large-scale self-hosted inference through vLLM, SGLang, or TokenSpeed
•Frontier-model routing experiments that compare task success, price, and review burden

⚙️Input Parameters

model

string

Use kimi-k3 on the Kimi API, k3 or k3-256k in Kimi Code, moonshotai/Kimi-K3 for the official checkpoint, or the exact model ID published by another provider.

messages

array

OpenAI-compatible conversation payload for text, multimodal, coding, reasoning, and tool-use workflows.

tools

array

Optional tool definitions for coding agents and other tool-using workflows where the active API surface supports them.

reasoning_effort

string

Optional reasoning control: low, high, or max. The Kimi API guide documents max as the default for kimi-k3, while Kimi Code documents high as the default for k3 and k3-256k.

💡Usage Examples

Example 1

Input Parameters

{
  "model": "kimi-k3",
  "messages": [
    {
      "role": "user",
      "content": "Inspect this repository plan, identify the riskiest assumption, and propose the smallest verifiable implementation sequence."
    }
  ]
}

Output Results

Kimi K3 returns a structured coding or reasoning response through the Kimi API. Validate the exact response shape, effort controls, tool support, and account limits against the current API documentation before production use.

Quick Actions

Use NowView Documentation

Technical Specifications

Hardware Type: Kimi API / Kimi Code k3 or k3-256k / Kimi Work / self-hosted 64+ accelerator supernode via vLLM, SGLang, or TokenSpeed
Commercial Use: Supported
Pricing: Kimi international API: $0.30 per million cache-hit input tokens, $3 per million cache-miss input tokens, and $15 per million output tokens; self-hosting is governed by the Kimi K3 License
Platform: Replicate

Related Keywords

Kimi K3kimi-k3Moonshot AI Kimi K3Kimi K3 APIKimi K3 pricingKimi K3 context windowKimi K3 open weightsKimi K3 coding modelKimi K3 vision modelKimi K3 256Kk3-256kKimi Delta Attention

Related Models

Bielik 1.5B v3 Instruct

Bielik-1.5B-v3-Instruct is a generative text model featuring 1.6 billion parameters. It is result of collaboration between the open-science/open-souce project SpeakLeash and the High Performance Computing (HPC)

Cordia-A6 Text Generation Model

A model for generating text sequences based on input prompts and adjustable parameters.

Claude Sonnet 4

Claude Sonnet 4 is a significant upgrade to 3.7, delivering superior coding and reasoning while responding more precisely to your instructions