Kimi-K3

Model IntroductionText

Kimi K3 is an open-weight, native multimodal agentic model and our most capable model to date.

Introduction

Kimi K3 is an open-weight, native multimodal agentic model and our most capable model to date. It is a 2.8T-parameter model built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), with native vision capabilities and a 1-million-token context window. It is the world's first open 3T-class model, designed for frontier intelligence across long-horizon coding, knowledge work, and reasoning.

Key Features

  • New Architecture: Kimi K3 is built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), and scales up MoE sparsity with a Stable LatentMoE framework that activates 16 out of 896 experts — yielding an approximate 2.5× improvement in overall scaling efficiency over Kimi K2.
  • Long-Horizon Coding: Operating with minimal human oversight, Kimi K3 sustains long engineering sessions, navigates massive repositories, and orchestrates terminal tools — from GPU kernel optimization and compiler development to vision-in-the-loop game dev, CAD, and even chip design.
  • Agentic Knowledge Work: Kimi K3 advances end-to-end knowledge work, producing deep research with interactive visualizations, widgets and dashboards, and motion design and video editing, powered by its native multimodal architecture.
  • Native Multimodality & Long Context: Kimi K3 understands text, images, and video within the same model, and supports a 1-million-token context window.
  • Open Frontier Weights: We release the full Kimi K3 model weights under the Kimi K3 License, making frontier intelligence openly available for research, deployment, and further innovation.

Benchmarks

Benchmark (Area)Score*
GPQA Diamond (Reasoning & Knowledge)93.5 (Eval)
CritPt (Reasoning & Knowledge)23.4 (Eval)
AA-LCR (Reasoning & Knowledge)74.7 (Eval)
HLE-Full (Reasoning & Knowledge)43.5 / 56.0 (Eval)
DeepSWE (Coding)67.5 (Eval)
ProgramBench (Coding)77.8 (Eval)
Terminal-Bench 2.1 (Coding)88.3 (Eval)
FrontierSWE (Coding)81.2 (Eval)
SWE-Marathon (Coding)42.0 (Eval)
PostTrainBench (Coding)36.6 (Eval)
MLS-Bench-Lite (Coding)48.3 (Eval)
SciCode (Coding)58.7 (Eval)
Kimi Code Bench 2.0 (Coding)72.9 (Eval)
BrowseComp (Agentic)91.2 (Eval)
DeepSearchQA (F1) (Agentic)95.0 (Eval)
ResearchRubrics (Agentic)76.2 (Eval)
GDPval-AA v2 (Elo) (Agentic)1686 (Eval)
Toolathlon-Verified (Agentic)76.5 (Eval)
MCPMark-Verified (Agentic)94.5 (Eval)
MCP-Atlas (Agentic)84.2 (Eval)
AutomationBench (Agentic)30.8 (Eval)
JobBench (Agentic)54.3 (Eval)
AA-Briefcase (Elo) (Agentic)1548 (Eval)
Agents' Last Exam (Agentic)28.3 (Eval)
APEX-Agents (Agentic)41.0 (Eval)
OfficeQA Pro (Agentic)63.3 (Eval)
SpreadsheetBench 2 (Agentic)34.8 (Eval)
OSWorld-Verified (Agentic)84.8 (Eval)
OSWorld 2.0 (Agentic)58.3 (Eval)
SaaS-Bench (Agentic)60.1 (Eval)
τ³-Banking (Agentic)33.4 (Eval)
Harvey Lab-AA (Agentic)94.6 (Eval)
CorpFin v2 (Agentic)71.6 (Eval)
Finance Agent v2 (Agentic)54.4 (Eval)
Legal Research Bench (Agentic)44.2 (Eval)
WorldVQA ForceAnswer (Vision)51.0 (Eval)
OmniDocBench (Vision)91.1 (Eval)
PerceptionBench (Vision)58.5 (Eval)
Video-MME (w. sub) (Vision)90.0 (Eval)
MMVU (Vision)82.1 (Eval)
BabyVision w/ python (Vision)85.7 (Eval)
MMMU-Pro (Vision)81.6 / 83.4 (Eval)
CharXiv (RQ) (Vision)84.8 / 91.3 (Eval)
MathVision (Vision)94.3 / 97.8 (Eval)
ZeroBench (pass@5) (Vision)23.0 / 41.0 (Eval)

API request example

Endpoint /v1/chat/completions. Replace <API_KEY> with a key from the console.

Standard OpenAI Chat Completions. Swap `model` for the target slug.

curl https://api.ravenrouter.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <API_KEY>" \
  -d '{
  "model": "Kimi-K3",
  "messages": [
    {
      "role": "user",
      "content": "Hello! Introduce yourself in one sentence."
    }
  ],
  "max_tokens": 64
}'

More models: API Reference → Examples by model

Kimi-K3 API — $3.21/1M input, $16.07/1M output | RavenRouter AI