Official page: https://huggingface.co/moonshotai/Kimi-K3…

This is an open-weight model released by Moonshot AI with 2.8 trillion parameters (MoE architecture, ~104B activated), supporting native multimodality (text + images) and a 1 million token context window.

Weights are provided in Safetensors and other formats, under the Kimi K3 License.

The open weights of Kimi K3 were officially released on July 27, 2026, and can be viewed and downloaded primarily on Hugging Face.

moonshotai/Kimi-K3 · Hugging Face

Source: https://huggingface.co/moonshotai/Kimi-K3 Kimi K3

Chat (https://www.kimi.com/)Homepage (https://www.moonshot.ai/)

Hugging Face (https://huggingface.co/moonshotai)Twitter Follow (https://twitter.com/kimi_moonshot)Discord (https://discord.gg/TYU2fdJykW)ModelScope) (https://modelscope.cn/organization/moonshotai)

License (https://huggingface.co/moonshotai/Kimi-K3/blob/main/LICENSE)

📰Tech Blog (https://www.kimi.com/blog/kimi-k3)|📄Full Report (https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf)

https://huggingface.co/moonshotai/Kimi-K3#1-model-introduction1. Model Introduction

Kimi K3 is an open-weight, native multimodal agentic model and our most capable model to date. It is a 2.8T-parameter model built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), with native vision capabilities and a 1-million-token context window. It is the world’s first open 3T-class model, designed for frontier intelligence across long-horizon coding, knowledge work, and reasoning.

https://huggingface.co/moonshotai/Kimi-K3#key-featuresKey Features

• New Architecture: Kimi K3 is built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), and scales up MoE sparsity with a Stable LatentMoE framework that activates 16 out of 896 experts — yielding an approximate 2.5× improvement in overall scaling efficiency over Kimi K2.
• Long-Horizon Coding: Operating with minimal human oversight, Kimi K3 sustains long engineering sessions, navigates massive repositories, and orchestrates terminal tools — from GPU kernel optimization and compiler development to vision-in-the-loop game dev, CAD, and even chip design.
• Agentic Knowledge Work: Kimi K3 advances end-to-end knowledge work, producing deep research with interactive visualizations, widgets and dashboards, and motion design and video editing, powered by its native multimodal architecture.
• Native Multimodality & Long Context: Kimi K3 understands text, images, and video within the same model, and supports a 1-million-token context window.
• Open Frontier Weights: We release the full Kimi K3 model weights under the Kimi K3 License, making frontier intelligence openly available for research, deployment, and further innovation.

https://huggingface.co/moonshotai/Kimi-K3#2-model-summary2. Model Summary

ArchitectureMixture-of-Experts (MoE)Total Parameters2.8TActivated Parameters104BNumber of Layers93Number of Dense Layers1Attention-Layer Composition69 KDA + 24 Gated MLAAttention Hidden Dimension7168Number of Attention Heads96Latent MoE Dimension3584MoE Hidden Dimension(per Expert)3072Number of Experts896Selected Experts per Token16Number of Shared Experts2Vocabulary Size160KContext Length1048576Attention MechanismKDA & Gated MLAActivation FunctionSiTU-GLUVision EncoderMoonViT-V2Parameters of Vision Encoder401MQuantizationMXFP4 weights / MXFP8 activations (quantization-aware training)ModalityText, Image

https://huggingface.co/moonshotai/Kimi-K3#3-evaluation-results3. Evaluation Results

BenchmarkKimi K3 (max)Claude Fable 5 (max, w/ fallback)GPT-5.6 Sol (max)Claude Opus 4.8 (max)GPT-5.5 (xhigh)GLM-5.2 (max)Reasoning & KnowledgeGPQA Diamond93.592.694.191.093.591.2CritPt23.428.632.320.927.120.9AA-LCR74.770.073.767.774.371.3HLE-Full43.5 / 56.053.3 / 63.044.5 / 58.049.8 / 57.941.4 / 52.2—CodingDeepSWE67.570.073.059.067.046.2ProgramBench77.876.877.671.970.863.7Terminal-Bench 2.188.388.088.884.683.482.7FrontierSWE81.286.671.366.764.967.3SWE-Marathon42.035.039.040.014.013.0PostTrainBench36.641.434.634.128.434.3MLS-Bench-Lite48.349.946.242.835.540.4SciCode58.760.256.153.556.150.5Kimi Code Bench 2.072.976.964.871.769.064.2AgenticBrowseComp91.288.090.484.384.4—DeepSearchQA (F1)95.094.2—93.1——ResearchRubrics76.2—73.873.564.071.1GDPval-AA v2 (Elo)168617471736159314911510Toolathlon-Verified76.577.974.976.273.559.9MCPMark-Verified94.587.492.976.492.9—MCP-Atlas84.284.783.683.682.882.6AutomationBench30.829.129.727.222.712.9JobBench54.357.445.448.438.343.4AA-Briefcase (Elo)154815831495135411581260Agents’ Last Exam28.325.7†29.627.026.620.4APEX-Agents41.043.339.939.438.535.6OfficeQA Pro63.369.963.263.960.941.4SpreadsheetBench 234.834.732.431.629.128.1OSWorld-Verified84.885.083.083.479.0—OSWorld 2.058.366.162.655.749.5—SaaS-Bench60.1—61.456.143.8—τ3-Banking33.426.833.027.631.326.8Harvey Lab-AA94.693.687.291.186.391.0CorpFin v271.671.864.466.768.466.1Finance Agent v254.456.353.853.951.849.7Legal Research Bench44.249.548.143.840.431.3VisionWorldVQA ForceAnswer51.056.741.839.138.5—OmniDocBench91.189.885.887.989.4—PerceptionBench58.557.259.747.255.8—Video-MME (w. sub)90.0—89.586.089.3—MMVU82.1—81.279.281.7—BabyVision w/ python85.790.588.981.283.6—MMMU-Pro81.6 / 83.481.2 / 86.583.0 / 84.678.9 / 82.781.2 / 83.2—CharXiv (RQ)84.8 / 91.388.9 / 93.584.6 / 89.180.5 / 89.984.1 / 89.0—MathVision94.3 / 97.894.8 / 98.695.8 / 97.886.7 / 97.192.2 / 96.8—ZeroBench (pass@5)23.0 / 41.023.0 / 46.017.0 / 35.017.0 / 34.022.0 / 41.0— FootnotesAll Kimi K3 results are obtained with reasoning effort set to ‘max’ and temperature = 1.0. For single-step tasks, such as GPQA Diamond, HLE-Full, and vision benchmarks without tools, we set top-p = 0.95; for agentic tasks, we set top-p = 1.0. For HLE-Full, MMMU-Pro, CharXiv (RQ), MathVision, and ZeroBench, each cell reports the scores without and with tool augmentation (general tools for HLE-Full, Python for the vision benchmarks), in that order.

1. Reasoning & knowledge benchmarks- **CritPt and AA-LCR.**Scores are cited fromArtificial Analysis (https://artificialanalysis.ai/)as of July 23, 2026.
2. Coding benchmarks- **DeepSWE.**Kimi K3 is evaluated with the Kimi Code harness. The GLM-5.2 score is taken from theGLM-5.2 release blog (https://z.ai/blog/glm-5.2); all remaining scores are from the officialDeepSWE leaderboard (https://deepswe.datacurve.ai/), under which Kimi K3 attains 67.3 with the mini-SWE-agent harness. We report the DeepSWE v1.1 tasks. - **Terminal-Bench 2.1.**Kimi K3 is evaluated with the Kimi Code harness. For all other models, we report the best score across harnesses: GLM-5.2 with Claude Code (GLM-5.2 release blog (https://z.ai/blog/glm-5.2)); Claude Opus 4.8 and Claude Fable 5 with Terminus 2 (Artificial Analysis (https://artificialanalysis.ai/evaluations/terminalbench-v2-1)); GPT-5.5 and GPT-5.6 Sol with Codex (OpenAI (https://openai.com/index/previewing-gpt-5-6-sol/)). - **ProgramBench.**Kimi K3 is evaluated with the Kimi Code harness. The GLM-5.2 score is from theGLM-5.2 release blog (https://z.ai/blog/glm-5.2); all other scores are fromVals AI (https://www.vals.ai/benchmarks/programbench). - **SWE-Marathon.**Kimi K3, Claude Opus 4.8, and Claude Fable 5 are evaluated with the Claude Code harness; GPT-5.6 Sol is evaluated with the Codex harness. The GLM-5.2 score is from theGLM-5.2 release blog (https://z.ai/blog/glm-5.2). Our evaluation is based on an H20-calibrated branch of theofficial tasks (https://www.swe-marathon.org/)as of July 9, 2026, prior to the final v1.1 release: the Docker images, performance gates, and reference oracles for the GPU tasks have been recalibrated for H20, while the correctness and anti-cheat validators remain unchanged. Additionally, Claude Fable 5 hit fallbacks on 35% of the tasks in our evaluation, which may have negatively impacted its measured performance. - **FrontierSWE.**Kimi K3 is evaluated with the Kimi Code harness and GPT-5.6 Sol with the Codex harness; all other results are fromFrontierSWE (https://www.frontierswe.com/). Dominance scores are recomputed from the raw scores using the official evaluation script and are current as of July 16, 2026. - **PostTrainBench.**Scores for GLM-5.2, GPT-5.5, and Claude Opus 4.8 are adopted from the officialPostTrainBench (https://posttrainbench.com/)results. Kimi K3, Claude Fable 5, and GPT-5.6 Sol are evaluated with the official Harbor implementation at maximum reasoning effort, averaged over three runs on H20 GPUs (instead of H100 in the official setting) — Kimi K3 and Claude Fable 5 with the Claude Code harness, and GPT-5.6 Sol with the Codex harness. - **MLS-Bench-Lite.**Kimi K3 is evaluated with the Kimi Code harness; GLM-5.2 and the Claude models with the Claude Code harness; GPT-5.5 and GPT-5.6 Sol with the Codex harness. - **SciCode.**Scores are cited fromArtificial Analysis (https://artificialanalysis.ai/)as of July 23, 2026. - **Kimi Code Bench 2.0 (in-house).**Kimi K3 is evaluated with the Kimi Code harness (it attains 73.7 with the Claude Code harness); GLM-5.2, Claude Opus 4.8, and Claude Fable 5 with the Claude Code harness; GPT-5.5 and GPT-5.6 Sol with the Codex harness. All models are evaluated at maximum reasoning effort, except GPT-5.5, which uses the “xhigh” setting. As the benchmark includes cybersecurity and safety-related tasks, we also disclose the fraction of refused or fallback tasks: Claude Fable 5 hit 13 fallbacks and 1 refusal out of 80 tasks; 10 refusals out of 80 tasks entered GPT-5.6 Sol’s cyber guard; GPT-5.5 had 3 refusals out of 80 tasks.
3. Agentic benchmarks- **OfficeQA Pro.**Each test case provides the agent with the entire PDF corpus, with all PDFs rendered as images and no machine-readable text available. - **OfficeQA Pro and SpreadsheetBench 2.**Kimi K3, GLM-5.2, Claude Opus 4.8, and Claude Fable 5 are evaluated with the Claude Code harness; GPT-5.5 and GPT-5.6 Sol are evaluated with the Codex harness. - **MCP-Atlas.**All models are evaluated on the 500-task public subset with a 100-turn limit, using Gemini 3.1 Pro as the judge. - **AutomationBench.**All models are evaluated on the 600-task public subset, following the official GitHub setup in all other respects. - **BrowseComp.**We adopt a context-compaction strategy triggered at 300K tokens. When evaluated with the full 1M-token context window and no context management, Kimi K3 achieves a score of 90.4. The results of Claude Fable 5, Claude Opus 4.8, GPT-5.6 Sol, and GPT-5.5 are cited fromAnthropic (https://www.anthropic.com/news/claude-fable-5-mythos-5)andOpenAI (https://openai.com/index/gpt-5-6/). - **GDPval-AA v2, AA-Briefcase, τ3-Banking, Harvey Lab-AA, and APEX-Agents.**Scores are cited fromArtificial Analysis (https://artificialanalysis.ai/)and theAPEX-Agents leaderboard (https://www.mercor.com/apex/apex-agents-leaderboard/)as of July 23, 2026. For Harvey Lab-AA, we report the criterion pass rate. - **CorpFin v2, Finance Agent v2, and Legal Research Bench.**Scores are cited fromVals AI (https://www.vals.ai/). - **Agents’ Last Exam.**Scores are cited from theofficial leaderboard (https://agents-last-exam.org/leaderboard)as of July 23, 2026; we report the leaderboard’s primary pass-rate metric. On the leaderboard, each model is paired with a specific harness: Kimi K3 with Kimi Code; GPT-5.6 Sol and GPT-5.5 with Codex; Claude Fable 5, Claude Opus 4.8, and GLM-5.2 with Claude Code.†The Claude Fable 5 entry runs at xhigh effort with 40% of tasks annotated as downgraded.
4. Multimodal benchmarks- Except for ZeroBench, which follows the official setting and is run five times, all multimodal scores are averaged over three runs. MMMU-Pro is evaluated following the official protocol, preserving the original input order and prepending images to the text input. - PerceptionBenchis an in-house benchmark that focuses on atomic visual perception capabilities.

https://huggingface.co/moonshotai/Kimi-K3#4-native-mxfp4-quantization4. Native MXFP4 Quantization

Kimi K3 applies quantization-aware training from the SFT stage onward, using MXFP4 weights with MXFP8 activations for broad hardware compatibility.

https://huggingface.co/moonshotai/Kimi-K3#5-deployment5. Deployment

You can access Kimi K3’s API onhttps://platform.kimi.ai/by selectingkimi\-k3, and we provide OpenAI/Anthropic-compatible API for you. Currently, Kimi K3 is recommended to run on the following inference engines:

• vLLM (https://github.com/vllm-project/vllm)— seerecipes (https://recipes.vllm.ai/moonshotai/Kimi-K3)
• SGLang (https://github.com/sgl-project/sglang)— seecookbook (https://docs.sglang.io/cookbook/autoregressive/Moonshotai/Kimi-K3)
• TokenSpeed (https://github.com/lightseekorg/tokenspeed)— seerecipes (https://lightseek.org/tokenspeed/recipes/models#kimi-k3)

https://huggingface.co/moonshotai/Kimi-K3#6-model-usage6. Model Usage

Kimi K3 always has thinking enabled, and will returnreasoning\_content. Thinking effort is configured with the top-levelreasoning\_effortrequest field, which supports"low","high", and"max"(default"max").

Kimi K3 was trained in the preserved thinking history mode. For multi-turn conversations and tool calls, Kimi K3 requires the complete assistant message returned by the API to be passed back tomessagesas-is — includingreasoning\_contentandtool\_calls, not justcontent:

`` import openai

def chat_with_preserved_thinking(client: openai.OpenAI, model_name: str): messages = [ { “role”: “user”, “content”: “Tell me three random numbers.” }, { “role”: “assistant”, “reasoning_content”: “I’ll start by listing five numbers: 473, 921, 235, 215, 222, and I’ll tell you the first three.”, “content”: “473, 921, 235” }, { “role”: “user”, “content”: “What are the other two numbers you have in mind?” } ]

response = client.chat.completions.create(
    model=model_name,
    messages=messages,
    stream=False,
    max_tokens=4096,
    reasoning_effort="max",
)
# the assistant should mention 215 and 222 that appear in the prior reasoning content
print(f"response: {response.choices[0].message.reasoning}")
return response.choices[0].message.content

``

For full guides and examples (vision input, structured output, partial mode, tool choice, dynamic tool loading, context caching), see theKimi K3 Quickstart (https://platform.kimi.ai/docs/guide/kimi-k3-quickstart)andThinking Effort (https://platform.kimi.ai/docs/guide/use-thinking-effort).

https://huggingface.co/moonshotai/Kimi-K3#coding-agent-frameworkCoding Agent Framework

Kimi K3 works best withKimi Code CLI (https://www.kimi.com/code)as its agent framework. We warmly invite you to give it a try — run Kimi Code in your terminal and select Kimi K3 using the/modelcommand. We hope you enjoy building with Kimi K3, and we would love to hear your feedback!

https://huggingface.co/moonshotai/Kimi-K3#7-license7. License

Both the code repository and the model weights are released under theKimi K3 License (https://huggingface.co/moonshotai/Kimi-K3/blob/main/LICENSE).

https://huggingface.co/moonshotai/Kimi-K3#8-contact-us8. Contact Us

If you have any questions, please reach out at[email protected].

Downloads last month2,850

Model tree formoonshotai/Kimi-K3https://huggingface.co/docs/hub/model-cards#specifying-a-base-model

Spaces usin