Model introduction
Kimi K3 is Moonshot AI's 2.8T-parameter native multimodal reasoning model with a 1M context window, built for long-horizon coding, knowledge work, and deep reasoning.
Model capabilities
Specifications come from the cited model material. Catalog tags help identify access features; use the source documentation for exact limits.
Verified specifications
- Official positioning
- Long-horizon coding, knowledge work, and deep reasoning
- Context window
- 1,000,000 tokens
- Model size
- 2.8T parameters, 16 / 896 experts active
- Input modalities
- Text, image
Sources: Moonshot AI Kimi K3 technical blog
Independent capability evaluation
Only public measurements matched to the exact model ID and reasoning profile are shown; nearby variants are not substituted.
Intelligence index
57
Cohort rank #4 / 186
Output speed
35.2 tok/s
Cohort rank #140 / 186
First-token latency
4.54s
Reasoning, max effort
A very high composite score with slow output and heavy token use, better for valuable long tasks than low-latency interaction.
Evaluation profile: Reasoning, max effort. Artificial Analysis · Methodology
Model pricing and access
Prices are read directly from the MoleAPI console catalog and shown by current billing group and context tier.
Your final charge follows the account group shown in the console.
Live pricing source
Pricing table
USD / 1M tokensStandardx1defaultInput $3 · Output $15
Input
$3 / 1M tokens
Output
$15 / 1M tokens
Cache read
$0.3 / 1M tokens
Discountx0.8discountInput $2.4 · Output $12
Input
$2.4 / 1M tokens
Output
$12 / 1M tokens
Cache read
$0.24 / 1M tokens
Relayx0.3relayInput $0.9 · Output $4.5
Input
$0.9 / 1M tokens
Output
$4.5 / 1M tokens
Cache read
$0.09 / 1M tokens
Temporary (availability may vary)x0.1tempInput $0.3 · Output $1.5
Input
$0.3 / 1M tokens
Output
$1.5 / 1M tokens
Cache read
$0.03 / 1M tokens
Access protocols
Available billing groups: Standard, Discount, Relay, Temporary (availability may vary)
- openai
- POST
- /v1/chat/completions
- openai-response
- POST
- /v1/responses
- anthropic
- POST
- /v1/messages
- gemini
- POST
- /v1beta/models/{model}:generateContent
Complete model guide
Prepared from vendor sources and independently reviewed. · 2026-07-28
Kimi K3 Model Overview
Kimi K3 is Kimi's flagship model for long-horizon programming, knowledge work, and reasoning. The vendor describes it as an open 3T-class model with native vision capabilities and a million-token context window. [1]
Model Positioning and Core Capabilities
Kimi K3 is designed for complex tasks that require sustained long-context handling, multi-step reasoning, and tool collaboration, covering code engineering, research analysis, and vision-enabled development workflows. [1]
The model supports text and image input, text output, and provides reasoning, tool calling, JSON Mode, and structured output capabilities. [7] Officially demonstrated use cases include sustained development in large codebases, terminal tool orchestration, frontend and game development with screenshots, CAD-related work, and literature retrieval with interactive report generation. [1]
Capabilities and Specifications
Kimi K3 uses a sparse MoE architecture with 2.8 trillion total parameters, activating 16 of 896 experts per inference. [1] Its architecture includes Kimi Delta Attention, Attention Residuals, and Stable LatentMoE. [1]
| Specification | Details |
|---|---|
| Model size | 2.8 trillion parameters. [1] |
| Architecture | Kimi Delta Attention, Attention Residuals, Stable LatentMoE. [1] |
| Expert configuration | Activates 16 of 896 experts. [1] |
| Context window | 1,048,576 tokens. [7] |
| Input modalities | Text, images. [7] |
| Output modalities | Text. [7] |
| Reasoning and output controls | Supports reasoning effort settings, tool calling, JSON Mode, and structured output. [7] |
According to the official statement, Kimi K3 delivers approximately 2.5× higher overall scaling efficiency than Kimi K2. [1]
Use Cases
Kimi K3's long context and sparse expert architecture target large-scale engineering codebases, long-horizon knowledge tasks, and multi-stage reasoning workflows. [1]
For programming, the model covers large codebase maintenance, tool use in terminal environments, GPU kernel optimization, compiler development, and chip design. [1] Its native vision capabilities can be used for frontend, game development, and CAD workflows that incorporate interface screenshots or visual assets. [1] For knowledge work, the vendor uses it for continuous tasks spanning information retrieval, analysis, and interactive visual report generation. [1]
Important Limitations
At launch, Kimi K3 defaults to its highest reasoning effort setting, while lower and higher reasoning effort modes are listed as future updates. [1]
The vendor positions Kimi K3 as an open model, but plans to release the full model weights by July 27, 2026. [1] The official code, research, and visual development examples are vendor demonstrations and do not constitute independent benchmark conclusions. [1]
Sources
[1] Moonshot AI Kimi K3 技术博客 [7] Kimi K3 Pricing | Plans, Membership & API Costs
Code examples
These examples use MoleAPI's Responses API endpoint and run after you replace the API key.
curl https://api.moleapi.com/v1/responses \
-H "Authorization: Bearer $MOLEAPI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "kimi-k3",
"input": "Explain this problem step by step."
}'Frequently asked questions
How is kimi-k3 priced?
This page reads prices from the MoleAPI console API and updates with model prices, context tiers, and account groups.
How can I access kimi-k3?
Protocols and endpoints come from the supported_endpoint_types field in the MoleAPI model catalog.
How do I switch an existing project to kimi-k3?
Keep the MoleAPI API address and key, replace the model parameter with the model ID on this page, then check protocol-specific parameter differences.
Where do the kimi-k3 model details come from?
Capabilities and limitations are checked against Moonshot AI and the other cited pages. Pricing and available protocols come only from the MoleAPI console.