Kimi-K3 is an open-weight, native multimodal agentic model developed by Moonshot AI for long-horizon coding, knowledge work, visual understanding, and reasoning. It is a 2.8T-parameter Mixture-of-Experts model built with Kimi Delta Attention, Attention Residuals, and Stable LatentMoE, with 104B activated parameters and a 1M-token context window.
This model is ready for commercial or non-commercial use.
This model is not owned or developed by NVIDIA. This model has been developed and built to a third-party's requirements for this application and use case; see link to Non-NVIDIA Kimi-K3 Model Card.
GOVERNING TERMS: This trial service is governed by the NVIDIA API Trial Terms of Service. Use of this model is governed by the NVIDIA Open Model Agreement. Additional Information: Modified MIT License. Kimi K3.
Global
Use Case: Developers and researchers can use Kimi-K3 for long-horizon software engineering, agentic knowledge work, multimodal document understanding, reasoning, tool use, visual content analysis, and interactive application development.
Build.NVIDIA.com: 08/20/2026 via link
Huggingface: 07/16/2026 via link
References:
Architecture Type: Transformer
Network Architecture: Mixture-of-Experts
Total Parameters: 2.8T
Active Parameters: 104B
Vocabulary Size: 160K
Input Context Length (ISL): 1,048,576 tokens
Base Model: Kimi-K3
Input Types: Text, Image
Input Formats: String, Red, Green, Blue (RGB)
Input Parameters: One-Dimensional (1D), Two-Dimensional (2D)
Other Input Properties: Supports multimodal conversations, system prompts, tool definitions, and preserved reasoning history. For multi-turn conversations and tool calls, clients must pass back the complete assistant message, including reasoning_content and tool_calls.
Output Types: Text
Output Format: String
Output Parameters: One-Dimensional (1D)
Other Output Properties: Supports reasoning content, structured output, function and tool calls, and configurable low, high, or max reasoning effort. Thinking is always enabled.
Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA's hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.
Runtime Engines:
Supported Hardware: NVIDIA Blackwell
Operating Systems: Linux
The integration of foundation and fine-tuned models into AI systems requires additional testing using use-case-specific data to ensure safe and effective deployment. Following the V-model methodology, iterative testing and validation at both unit and system levels are essential to mitigate risks, meet technical and functional requirements, and ensure compliance with safety and ethical standards before deployment.
Kimi-K3
Data Modality: Text, Image
Image Training Data Size: Undisclosed
Text Training Data Size: Undisclosed
Training Data Collection: Undisclosed
Training Labeling: Undisclosed
Training Properties: Kimi-K3 uses quantization-aware training from the supervised fine-tuning stage onward with MXFP4 weights and MXFP8 activations. Specific pretraining datasets, collection methods, and labeling methods are not disclosed in the source Model Card.
Testing Data Collection: Undisclosed
Testing Labeling: Undisclosed
Testing Properties: Undisclosed
Evaluation Benchmark Score: Kimi-K3 was evaluated across reasoning and knowledge, coding, agentic, and multimodal benchmarks. Selected partner-reported results are shown below.
| Category | Benchmark | Kimi-K3 Result |
|---|---|---|
| Reasoning and Knowledge | GPQA Diamond | 93.5 |
| Reasoning and Knowledge | AA-LCR | 74.7 |
| Coding | DeepSWE | 67.5 |
| Coding | ProgramBench | 77.8 |
| Coding | Terminal-Bench 2.1 | 88.3 |
| Coding | FrontierSWE | 81.2 |
| Multimodal | MMVU | 82.1 |
| Multimodal | BabyVision with Python | 85.7 |
| Multimodal | MMMU-Pro, without/with tools | 81.6 / 83.4 |
| Multimodal | MathVision, without/with tools | 94.3 / 97.8 |
Evaluation Data Collection: Hybrid: Automated, Manually-Collected
Evaluation Labeling: Hybrid: Automated, Manually-Labeled
Evaluation Properties: Partner-reported Kimi-K3 results use max reasoning effort and temperature 1.0. Single-step tasks use top-p 0.95, while agentic tasks use top-p 1.0. Tool-augmented benchmark cells report scores without and with tool augmentation in that order. Coding benchmarks use benchmark-specific Kimi Code, Claude Code, Codex, or official harnesses; multimodal results other than ZeroBench are averaged over three runs.
Acceleration Engine(s): vLLM; Dynamo
Test Hardware: NVIDIA Grace Blackwell GB300x8
Precision Formats: MXFP4 weights with MXFP8 activations
Kimi-K3 was trained with preserved thinking history. Applications using multi-turn conversations or tool calls must return the complete prior assistant message to the model, including reasoning content and tool calls. Refer to the Kimi-K3 Quickstart for current API guidance.
NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
Users are responsible for model inputs and outputs. Users are responsible for ensuring safe integration of this model, including implementing guardrails as well as other safety mechanisms, prior to deployment.
Please make sure you have proper rights and permissions for all input image content; if the image includes people, personal health information, or intellectual property, the image generated will not blur or maintain proportions of image subjects included.
Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns here