Kimi-K2.5-MXFP8

Model Overview

Description:

The NVIDIA Kimi-K2.5-MXFP8 model is a quantized version of Moonshot AI's Kimi-K2.5 model, a native multimodal agentic model with Mixture of Experts (MoE) architecture. Kimi-K2.5 has 1T total parameters with 32B activated parameters, 384 routed experts (8 selected per token), and 61 transformer layers. For more information, refer to the Kimi-K2.5 model card. The NVIDIA Kimi-K2.5-MXFP8 model was quantized using the TensorRT Model Optimizer.

This model is ready for commercial/non-commercial use. <br>

Third-Party Community Consideration

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-K2.5) Model Card.

License/Terms of Use:

Modified MIT

Deployment Geography:

Global <br>

Use Case:

Developers looking to take off the shelf, pre-quantized models for deployment in AI Agent systems, chatbots, RAG systems, multimodal applications, and other AI-powered applications. <br>

Release Date:

Huggingface via https://huggingface.co/vincentzed-hf/Kimi-K2.5-MXFP8 <br>

Model Architecture:

Architecture Type: Transformers (Mixture of Experts) <br> Network Architecture: KimiK25ForConditionalGeneration (DeepseekV3-based) <br> **This model was developed based on Kimi-K2.5 <br> Total Parameters: 1T <br> Activated Parameters: 32B <br> Number of Layers: 61 (including 1 dense layer) <br> Number of Experts: 384 routed, 1 shared, 8 selected per token <br> Vision Encoder: MoonViT (400M parameters) <br>

Input:

Input Type(s): Text, Image, Video <br> Input Format(s): String, Image tensors <br> Input Parameters: Multi-modal <br>

Output:

Output Type(s): Text <br> Output Format: String <br> Output Parameters: 1D (One-Dimensional): Sequences <br>

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. <br>

Software Integration:

Runtime Engine(s): <br>

• SGLang <br>

Supported Hardware Microarchitecture Compatibility: <br>

• NVIDIA Blackwell <br>

Preferred Operating System(s): <br>

• Linux <br>

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.

Model Version(s):

** The model is quantized with nvidia-modelopt 0.41.0rc2.dev72+g886781332 <br>

Training, Testing, and Evaluation Datasets:

Calibration Dataset:

• Link: Nemotron-Post-Training-Dataset-v2 <br>
• Data collection method: Automated. <br>
• Labeling method: Automated. <br>

Training Datasets:

• Data Collection Method by Dataset: Undisclosed <br>
• Labeling Method by Dataset: Undisclosed<br>
• Properties: Undisclosed

Testing Dataset:

• Data Collection Method by Dataset: Undisclosed <br>
• Labeling Method by Dataset: Undisclosed <br>
• Properties: Undisclosed <br>

Evaluation Dataset:

• Data collection method: Hybrid: Automated, Human <br>
• Labeling method: Hybrid: Human, Automated <br>

Inference:

Acceleration Engine: SGLang <br> Test Hardware: B300 <br>

Post Training Quantization

This model was obtained by quantizing the weights of Kimi-K2.5 to MXFP8 data type, ready for inference with SGLang. Only the weights of the linear operators within transformer blocks are quantized (excluding attention projections, vision tower, and mm_projector). This optimization reduces the number of bits per parameter from 16 to 8, reducing the disk size and GPU memory requirements by approximately 2x.

Usage

Deploy with SGLang

To serve the quantized MXFP8 checkpoint with SGLang:

python3 -m sglang.launch_server --model-path vincentzed-hf/Kimi-K2.5-MXFP8 --quantization modelopt

Please install from source: git clone [email protected]:sgl-project/sglang.git Once the repo is cloned, do uv pip install -e "python[all]" and run the serve command.

Reproduce with ModelOpt

You may want to produce this checkpoint yourself. To reproduce the MXFP8 quantized checkpoint using TensorRT Model Optimizer:

python3 examples/llm_ptq/hf_ptq.py \
    --pyt_ckpt_path /root/.cache/huggingface/hub/models--moonshotai--Kimi-K2.5/snapshots/c0d6821ed3d48201b834278fb99d8f2d37732a52 \
    --qformat mxfp8 \
    --kv_cache_qformat none \
    --export_path ./kimi-k2.5-mxfp8 \
    --trust_remote_code

Evaluation

The accuracy benchmark results will be updated:

<table> <tr> <td><strong>Precision</strong> </td> <td><strong>Benchmark 1</strong> </td> <td><strong>Benchmark 2</strong> </td> </tr> <tr> <td>BF16 </td> <td><!-- TODO --> </td> <td><!-- TODO --> </td> </tr> <tr> <td>MXFP8 </td> <td><!-- TODO --> </td> <td><!-- TODO --> </td> </tr> </table>

Baseline: Kimi-K2.5.

Model Limitations:

The base model was trained on data that contains toxic language and societal biases originally crawled from the internet. Therefore, the model may amplify those biases and return toxic responses especially when prompted with toxic prompts. The model may generate answers that may be inaccurate, omit key information, or include irrelevant or redundant text producing socially unacceptable or undesirable text, even if the prompt itself does not include anything explicitly offensive.

Ethical Considerations

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.

Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns here.