LLMPrivate

Kimi K3

Moonshot AI's 2.8T-parameter open-weight flagship with native vision, 1M context, and frontier coding capabilities.

Maker
Moonshot AI
Modality
Text + vision
Context
1,000K tokens
Open weights
Yes — released under the Kimi K3 License

Overview

What is Kimi K3

Kimi K3 is Moonshot AI's flagship open-weight multimodal language model released in July 2026. At 2.8 trillion parameters, it is the first open 3T-class model, built on Kimi Delta Attention and Attention Residuals, with native vision, a 1-million-token context window, and advanced reasoning for coding and agentic tasks.

Running it privately on Venice

On Venice, Kimi K3 runs under a private, zero-retention privacy tier — your prompts are not stored, profiled, or used for training. You get the full open-weight frontier model with vision, reasoning, tool use, and web search, pay per token without a subscription, and retain full data sovereignty.

Private (zero retention)No prompt trainingTEE · hardware enclaveEnd-to-end encrypted

Assessment

Strengths and limitations

Strengths
  • First open model in the 3-trillion-parameter class (2.8T), with downloadable weights for self-hosting and fine-tuning.
  • Native multimodal understanding of text, images, and video within a single model.
  • 1-million-token context window enables analysis of massive codebases, documents, and long-form video.
  • Leads on long-horizon coding benchmarks including SWE Marathon and Arena.ai's Code Arena WebDev.
  • Strong agentic stack: tool use, function calling, reasoning, web search, and structured JSON output.
  • Open-weight availability under the Kimi K3 License provides deployment flexibility without vendor lock-in.
Limitations
  • Overall general-intelligence performance trails top proprietary models Claude Fable 5 and GPT-5.6 Sol.
  • Reported higher hallucination rate than its predecessor Kimi K2.6.
  • Expensive at scale on Venice and computationally demanding to self-host due to the massive MoE architecture.
  • Not uncensored: content moderation policies apply to the model outputs.

Capabilities

What it supports

  • Tool use / function calling
  • Vision (image input)
  • Reasoning
  • Web search
  • Code-optimized
  • Structured output (JSON schema)
  • Audio input
  • Video input
  • Multiple image inputs
  • Log probabilities

Specifications

Datasheet

Maker
Moonshot AI
Released
July 16, 2026
Architecture
Mixture-of-Experts with Kimi Delta Attention (KDA) and Attention Residuals; 2.8T total parameters, 16 of 896 experts active per token
Parameters
2.8 trillion total
Modality
Native multimodal — text, image, and video input; text output
Open weights
Yes — released under the Kimi K3 License
Context window
1,000K tokens
Max output
131.072K tokens
Capabilities
Vision, Function calling, Reasoning, Web search, Code-optimized
Privacy on Venice
Private — zero retention
Available on Venice since
Jul 2026

API

Call it from your code

Venice exposes an OpenAI-compatible API. Point your base URL at Venice and pass the model id.

curl https://api.venice.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kimi-k3",
    "messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
  }'

Pricing

What it costs on Venice

Billed per token on Venice: $4.69 per 1M input tokens and $23.44 per 1M output tokens.

Input / 1M tokens
$4.69
Per 1M tokens
Output / 1M tokens
$23.44
Per 1M tokens
Cached input / 1M
$0.47
Per 1M tokens

New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.

Alternatives

How it compares

ModelContext windowOpen weightsPrice (Venice)Best for
Kimi K31M tokensYes$4.69 in · $23.44 out / 1MLong-horizon coding
Claude Fable 51M tokensNo$12 in · $60 out / 1MGeneral reasoning
GPT-5.6 Sol1M tokensNo$6.25 in · $37.50 out / 1MFrontier tasks
DeepSeek V3.2160K tokensYes$0.33 in · $0.48 out / 1MEfficient open

The largest open-weight model on Venice, leading on coding and agentic benchmarks.

Use cases

What it is good for

  1. 01Long-horizon software engineering, repository navigation, and autonomous coding sessions.
  2. 02Deep research with interactive visualizations, dashboards, and multimodal document analysis.
  3. 03Agentic workflows that combine web search, tool use, and reasoning for complex knowledge work.
  4. 04Vision-in-the-loop tasks such as CAD, chip design, and game development.
  5. 05Processing and reasoning over very long inputs like books, legal transcripts, and video sequences.

Prompting

Getting better results

Feed full repositories or lengthy documents into the 1M context window to exploit its long-context reasoning.

Use structured JSON schema output for reliable agentic pipelines and automated tool chains.

For coding tasks, reference specific file paths and existing code patterns to maximize SWE Marathon-level accuracy.

Leverage native vision by uploading diagrams, screenshots, or video frames alongside text prompts.

Version history

Kimi K2.6
2025

Predecessor with lower hallucination rates but smaller scale.

Kimi K3
2026-07

Current — 2.8T open-weight flagship with native multimodality.

FAQ

Frequently asked questions

Kimi K3 is Moonshot AI's flagship open-weight multimodal language model, released in July 2026. At 2.8 trillion parameters, it is the first open 3T-class model, featuring native vision, a 1-million-token context window, and advanced reasoning for coding and agentic tasks.

On Venice, Kimi K3 costs $4.69 per million input tokens and $23.44 per million output tokens, with cached input at $0.47 per million. You pay per token with no subscription required.

You can try Kimi K3 on Venice using free credits included with your account. Beyond the trial, usage is billed per token in credits.

Yes. Moonshot AI released the full Kimi K3 weights under the Kimi K3 License, making it the largest open-weight model available. You can download and self-host it, or run it privately via Venice without infrastructure setup.

Yes. Kimi K3 supports function calling, reasoning, web search, and structured JSON output on Venice, making it suitable for agentic workflows and complex knowledge work.

Claude Fable 5 leads on general intelligence and real-world task benchmarks, while Kimi K3 wins on long-horizon coding and leads all open models. Kimi K3 is also open-weight and significantly cheaper per token on Venice.

Kimi K3 supports up to 1 million tokens of context and can generate up to 131,072 tokens in a single response.

Yes. Kimi K3 is natively multimodal and accepts both image and video input alongside text, enabling vision-in-the-loop coding, design, and analysis.

Run Kimi K3 privately

No prompt logging. No data used for training.