Kimi K3Kimi K3
Kimi K3 is now live

Open Agentic Intelligence
Built for Real Reasoning

Kimi K3 Thinking is Moonshot AI's 2.8-trillion-parameter open-source flagship. It blends native vision, a 1M-token context window, and max-effort reasoning to plan, explore, and solve through hundreds of deliberate steps across coding, search, and complex problem-solving.

1M Context
2.8T Parameters
Open Weights

What is Kimi K3?

Kimi K3 is Moonshot AI's 2.8-trillion-parameter open-source flagship. Built on a sparse mixture-of-experts architecture with Kimi Delta Attention (KDA) hybrid attention, it delivers a 1M-token context, native vision, and max-effort reasoning in a single, openly licensed model.

2.8T

Total Parameters

A 2.8-trillion-parameter sparse MoE with 896 experts, activating just 16 per token for efficient, scalable inference.

1M

Context Tokens

Up to 1M tokens on the Allegretto tier and above—enough to reason over entire codebases and long documents at once.

KDA

Architecture

Kimi Delta Attention: a hybrid linear-attention mechanism with attention residuals for long-context efficiency.

Key Features of Kimi K3

Engineered around agentic intelligence and grounded reasoning, making it ideal for tasks that demand autonomous problem-solving, tool use, and long-context analysis.

Agentic Capabilities

Purpose-built for tool use, reasoning, and autonomous problem-solving. It explores multiple paths, calls external tools, and executes complex multi-step tasks end to end.

Native Vision

Visual understanding is built in from the ground up—not bolted on. Read charts, screenshots, documents, and diagrams alongside text in a single context.

Max-Effort Reasoning

Ships with reasoning_effort set to max. It plans, backtracks from dead ends, and verifies conclusions—delivering state-of-the-art results on coding and knowledge tasks.

Advanced Features

Discover the innovative features that set Kimi K3 apart from other AI models.

1

Trillion-Scale MoE

2.8T total parameters across 896 experts for comprehensive knowledge across domains.

2

KDA Hybrid Attention

Kimi Delta Attention delivers efficient long-context reasoning with attention residuals.

3

Agentic Intelligence

Designed for autonomous action, tool use, and complex multi-step problem-solving.

4

Open Weights (MIT)

Full model weights released under a Modified MIT license for community use and development.

How to Use Kimi K3

Three ways to access the power of Kimi K3, from self-deployment to instant cloud access.

01

Self-Deploy Kimi K3

Kimi K3 is fully open-source under a Modified MIT license. If you have compatible hardware and deployment capability, you can download the model weights from GitHub or HuggingFace and deploy it on your own infrastructure.

02

Use Kimi Official API

For developers who want to integrate Kimi K3 into their applications, the official API is available at platform.moonshot.ai with comprehensive documentation and SDKs, billed at $3 in / $15 out per 1M tokens.

03

Use on kimik3.com

The easiest option. No deployment or API integration needed. Access Kimi K3 along with DeepSeek, Claude, GPT, Gemini, Qwen, and Grok all in one platform. New users get 3 free credits.

New User Get 3 Free Credits

Simple Pricing

One-time payment. No subscriptions. No hidden fees.

Starter

A great fit for the average user.

$20$9.9USD

One-Time Pay

  • 400 credits, valid for 1 month
  • Access to Kimi K3 + all major AI models
  • 1M-token context & native vision
  • Multi-language support
  • Real-time conversation
Get Started
Most Popular

Pro

Optimized for power users.

$100$49.9USD

One-Time Pay

  • 3,000 credits, valid for 6 months
  • Everything in Starter, plus:
  • Max-effort reasoning enabled
  • Ultra-fast response speed
  • Early access to beta features
  • Lifetime updates
Get Pro

Ultimate

Ultimate AI for professionals.

$200$99.9USD

One-Time Pay

  • 6,000 credits, valid for 1 year
  • Everything in Pro, plus:
  • Priority compute & highest throughput
  • Extended credit validity
  • Early access to beta features
  • Lifetime updates
Get Ultimate

Kimi API is billed separately at $3 / 1M input tokens and $15 / 1M output tokens, with cached input at just $0.30 / 1M. Mooncake serving keeps coding cache hit rates above 90%, cutting real input cost roughly 4×.

Frequently Asked Questions

Everything you need to know about Kimi K3.

Kimi-K3-Base is the foundation model pre-trained on a large-scale corpus, suitable for continued training and fine-tuning. Kimi-K3-Instruct is the instruction-tuned version optimized for conversational tasks, tool use, and following user directions. For most users, Kimi-K3-Instruct is the recommended starting point.

Start Using Kimi K3 Today

Get 3 free credits to experience the power of open-source agentic intelligence. No credit card required.