Writingby Moonshot AI

Kimi K3

Name: Kimi K3
Author: Moonshot AI

Free

The world's largest open-source LLM — a 2.8-trillion-parameter frontier model for coding, reasoning, and long-horizon knowledge work

Kimi K3 is Moonshot AI's flagship open-weight large language model, released on July 16, 2026, and billed as the world's first open-source model in the 3-trillion-parameter class with roughly 2.8 trillion total parameters. Built on a highly sparse Mixture-of-Experts architecture (activating 16 of 896 experts per inference) plus two internally developed innovations — Kimi Delta Attention (a hybrid linear attention mechanism) and Attention Residuals — it delivers a 1-million-token context window, native visual understanding, and an always-on 'thinking mode' for reasoning. K3 is designed for long-horizon coding, agentic workflows, and end-to-end knowledge work (research, drafting, and slide generation). Independent testing by Artificial Analysis placed it near the top of open models, competitive with Claude Opus 4.8 and GPT-5.5 while still trailing the strongest proprietary systems. It is accessible free via the Kimi web/app, through an OpenAI-SDK-compatible API, and via open weights for self-hosting (full weights scheduled for release by July 27, 2026).

Who it's for

developersAI researchersenterprises seeking low-cost open-weight modelsagentic-workflow builderstechnical teams working with large codebases

Pricing · freemium

checked 1d ago

Plan	Price	Includes
Free (Web & App)	$0 /mo	Access to Kimi K3 via kimi.com and the Kimi app · 1M-token context window · Native vision and always-on thinking mode · Swarm parallel task execution and Goal autonomous multi-step objectives
API (Usage-based)	Free	$3 per million input tokens · $15 per million output tokens · Cache-hit input drops to ~$0.30 per million (about 90% discount) · Flat pay-as-you-go pricing across the full 1M-token context, no tiering by context length · OpenAI-SDK compatible via base URL swap
Open Weights (Self-host)	$0 /mo	Full model weights released openly (by July 27, 2026) · Free to self-host on your own cluster · Requires multi-GPU infrastructure — not feasible on consumer hardware · Customizable and inspectable for enterprise deployment

AI-researched pricing — verify on the official site before subscribing.

Use it for

— Long-horizon coding across entire repositories (reading, debugging, patching, and documenting)
— Building autonomous agents and multi-step task pipelines
— End-to-end knowledge work: research synthesis, drafting, and slide generation
— Analyzing book-length documents, legal archives, or large datasets in a single pass
— Multimodal tasks combining text, images, and video with native visual understanding
— Self-hosted, customizable deployment for cost-sensitive or privacy-conscious organizations

Get the most out of it

01Hand K3 whole jobs rather than small questions — it excels at long-horizon tasks like 'read this repo, find the bug, patch it, and write the changelog'
02Use the 1M-token context to feed entire project directories or document sets instead of building a separate retrieval pipeline
03Exploit prompt caching on the API — cache-hit input tokens cost roughly 90% less ($0.30 vs $3 per million), so structure prompts to reuse context
04Migrate existing OpenAI/Anthropic code easily by swapping the base URL to Moonshot's endpoint and setting the model name to 'kimi-k3'
05Note that reasoning_effort currently only supports the 'max' level (default); avoid using the older K2.x thinking parameter
06If you plan to self-host, budget for cluster-grade GPUs — the multi-trillion-parameter MoE is beyond consumer hardware, so the web app or API is the practical route for most users

Visit Kimi K3

In the news

Was this useful?

Plan

Price

Includes

Free (Web & App)

$0 /mo

Access to Kimi K3 via kimi.com and the Kimi app · 1M-token context window · Native vision and always-on thinking mode · Swarm parallel task execution and Goal autonomous multi-step objectives

API (Usage-based)

Free

$3 per million input tokens · $15 per million output tokens · Cache-hit input drops to ~$0.30 per million (about 90% discount) · Flat pay-as-you-go pricing across the full 1M-token context, no tiering by context length · OpenAI-SDK compatible via base URL swap

Open Weights (Self-host)

$0 /mo

Full model weights released openly (by July 27, 2026) · Free to self-host on your own cluster · Requires multi-GPU infrastructure — not feasible on consumer hardware · Customizable and inspectable for enterprise deployment

Use it for

— Long-horizon coding across entire repositories (reading, debugging, patching, and documenting)

— Building autonomous agents and multi-step task pipelines

— End-to-end knowledge work: research synthesis, drafting, and slide generation

— Analyzing book-length documents, legal archives, or large datasets in a single pass

— Multimodal tasks combining text, images, and video with native visual understanding

— Self-hosted, customizable deployment for cost-sensitive or privacy-conscious organizations

Get the most out of it

01Hand K3 whole jobs rather than small questions — it excels at long-horizon tasks like 'read this repo, find the bug, patch it, and write the changelog'

02Use the 1M-token context to feed entire project directories or document sets instead of building a separate retrieval pipeline

03Exploit prompt caching on the API — cache-hit input tokens cost roughly 90% less ($0.30 vs $3 per million), so structure prompts to reuse context

04Migrate existing OpenAI/Anthropic code easily by swapping the base URL to Moonshot's endpoint and setting the model name to 'kimi-k3'

05Note that reasoning_effort currently only supports the 'max' level (default); avoid using the older K2.x thinking parameter

06If you plan to self-host, budget for cluster-grade GPUs — the multi-trillion-parameter MoE is beyond consumer hardware, so the web app or API is the practical route for most users