Kimi K3

Open 3T-class model for long-horizon coding and knowledge work

Try now

read docs

About model

Kimi K3 is Moonshot AI's most capable model and the first open model at the 3-trillion-parameter class, with 2.8T total parameters activating 16 of 896 experts per token. It is built on Kimi Delta Attention and Attention Residuals, two architectural changes to how information flows across sequence length and model depth, with native vision and a 1-million-token context window. The model is designed for long-horizon work: sustaining extended engineering sessions across large repositories, carrying multi-step research and document tasks end to end, and reading screenshots, charts, and documents inside the same model. It runs at maximum thinking effort at launch, and weights are released under an open license. Available on Together AI.

‍

Total Parameters

2.8T

Sparse MoE activating 16 of 896 experts per token

Context Window

Entire repositories or thousands of pages in a single request

Vision Input

Native

Screenshots, charts, and documents processed by the same model, no separate pipeline

Model key capabilities

Long-Horizon Coding: Sustains extended engineering sessions with minimal oversight, navigating large repositories and orchestrating terminal tools
Agentic Knowledge Work: Carries research, analysis, and document production tasks end to end across long multi-step sessions
Native Vision: Reads screenshots, charts, and documents directly, and iterates between code and rendered output for frontend and game development
Production-Ready Infrastructure: 99.9% SLA, available on serverless and dedicated infrastructure

Performance benchmarks

Model	FrontierMath Tier 4	GPQA Diamond	HLE	SciCode	GDPval-AA	Terminal-Bench 2.1	Agent Arena	FrontierCode	DeepSWE
Kimi K3	39.0%	93.5%	44%	59%	59%	85%	+15.5pp	44.2%	69%
Related open-source models
Competitor closed-source models
Claude Fable 5	87.8%	92.6%	53%	60%	62%	85%		53.5%	70%
Claude Opus 5	73.2%	93.2%	53%	56%	68%	89%	+16.4pp	53.4%	74%
GPT-5.6 Sol	82.9%	94.1%	47%	56%	61%	88%		47.5%	73%
Grok 4.5	24.4%	93.1%	40%	54%	51%	82%	+4.2pp	42.4%	54%
GPT-5.6 Luna	61.0%	91.1%	37%	53%	54%	81%		39.8%	67%

API usage

cURL
Python
Typescript

Endpoint:

moonshotai/Kimi-K3

curl -X POST "https://api.together.xyz/v1/chat/completions" \
  -H "Authorization: Bearer $TOGETHER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "moonshotai/Kimi-K3",
    "messages": [
      {
        "role": "user",
        "content": "What are some fun things to do in New York?"
      }
    ]
}'

from together import Together

client = Together()

response = client.chat.completions.create(
  model="moonshotai/Kimi-K3",
  messages=[
    {
      "role": "user",
      "content": "What are some fun things to do in New York?"
    }
  ]
)
print(response.choices[0].message.content)

import Together from 'together-ai';
const together = new Together();

const completion = await together.chat.completions.create({
  model: 'moonshotai/Kimi-K3',
  messages: [
    {
      role: 'user',
      content: 'What are some fun things to do in New York?'
     }
  ],
});

console.log(completion.choices[0].message.content);

Model card
Architecture Overview:
• 2.8T total parameters with Stable LatentMoE routing, activating 16 of 896 experts per token
• Built on Kimi Delta Attention (KDA) for efficient attention scaling and Attention Residuals (AttnRes) for selective retrieval of representations across depth
• Gated MLA and Sigmoid Tanh Unit (SiTU) components for attention selectivity and activation control
• 1-million-token context window with native vision input
• Quantization-aware training from the SFT stage onward, with MXFP4 weights and MXFP8 activations

Training Methodology:
• Moonshot AI reports an approximate 2.5x improvement in overall scaling efficiency over Kimi K2 from the combined architecture, training, and data changes
• Quantile Balancing derives expert allocation directly from router-score quantiles, removing a sensitive balancing hyperparameter
• Per-Head Muon optimizes attention heads independently for more adaptive learning at scale
• Trained in preserved thinking history mode, retaining reasoning content across turns within a session

Performance Characteristics:
• Terminal and repository coding (Terminal-Bench 2.1, completing real tasks in a command-line environment): 88.3
• Long-horizon web research (BrowseComp, multi-step web search questions, with context management): 91.2
• Expert knowledge and science (GPQA-Diamond, graduate-level science questions): 93.5; Humanity's Last Exam: 43.5, or 56.0 with tools
• Tool orchestration (MCP Atlas, agent tasks across many tools): 84.2
• Vision (MMMU-Pro university-level multimodal reasoning: 81.6; OmniDocBench document parsing: 91.1; MathVision: 94.3)
• Knowledge work (GDPval-AA v2 Elo on Artificial Analysis, real occupational tasks): 1668
‍
Prompting
Together AI API Access:
• Access Kimi K3 via Together AI APIs using the endpoint moonshotai/Kimi-K3
• Authenticate using your Together AI API key in request headers
• The model is trained with preserved thinking history: return prior reasoning content across turns within a session for stable generation quality
‍• For applications that need tight behavioral boundaries, set explicit constraints in the system prompt
• Available on Together AI serverless and dedicated infrastructure
‍
Applications & use cases
Long-Horizon Software Engineering:
• Run extended agent sessions that navigate large repositories and terminal tools with minimal supervision
• Hold entire codebases in the 1M-token context for cross-file refactors and reviews
• Iterate on frontend and game development using screenshots of rendered output as feedback

Research & Knowledge Work:
• Produce end-to-end research reports, analyses, and presentations across long multi-step sessions
• Cross-reference large document sets, filings, and papers inside a single context window
• Parse charts, tables, and scanned documents directly through native vision input

Multimodal Agent Systems:
• Build agents that reason over text and images together without a separate vision pipeline
• Process screenshots and visual state as part of automated workflows
• Sustain multi-step agent loops that keep reasoning context across turns
‍

Related models

Model specifications

Model data

Model provider
Moonshot AI
Type
Chat
Reasoning
Vision
Features
JSON Mode
Function Calling
Intelligence
Very High
Deployment
Serverless
Dedicated
Provisioned Throughput
Endpoint
moonshotai/Kimi-K3
Parameters
2.8T
Context length
1M
Input price
$3.00 / 1M tokens
$0.30 (cached)/1M
Output price
$15.00 / 1M tokens
Input modalities
Text
Image
Output modalities
Text

Released
July 27, 2026
Quantization level
FP4
External link
Provider docs
Category
Chat

Run in Playground

Quickstart docs

Deploy model

Kimi K3

About model

API usage

Model card

Prompting

Applications & use cases