---

🚀 Quick Start with Docker (Recommended)

You can easily run this model using the DGX-Spark-llama.cpp-Bench inference engine. It's pre-configured for high-performance inference on NVIDIA hardware (especially Blackwell/DGX Spark).

1. Pull the Docker Image

docker pull ghcr.io/sowilow/dgx-spark-llama.cpp-bench:latest

2. Run the Inference Server

For detailed configuration and usage, visit the GitHub Repository.

---

gpt-oss-20b-DGX-Spark-GGUF

This repository provides GGUF quantized versions of OpenAI's gpt-oss-20b, optimized specifically for NVIDIA Blackwell (DGX Spark) architectures.

These models were converted and quantized using llama.cpp with support for the gpt_oss architecture.

Model Highlights

• Optimized for Blackwell: Specifically tuned for high-performance inference on NVIDIA DGX Spark (SM120/SM121).
• Flexible Quantization:
  • Q4_MXFP4: 4-bit Medium quantization (recommended for efficiency).
  • Q8_0: 8-bit quantization (recommended for maximum precision).
• MoE Architecture: 21B total parameters with 3.6B active parameters, leveraging Mixture-of-Experts for high efficiency.
• Long Context: Supports up to 131k context length.

Quantization Details

File	Quant Method	Bitrate	Size	Description
gpt-oss-20b-q4_mxfp4.gguf	Q4_MXFP4	4.5 bpw	~12 GB	Balanced performance and quality.
gpt-oss-20b-q8_0.gguf	Q8_0	8.5 bpw	~22 GB	Standard 8-bit quantization.

Quick Start (llama.cpp)

To run these models on a DGX Spark system:

1. Pull the optimized Docker image:

   docker pull ghcr.io/sowilow/dgx-spark-llama.cpp-bench:latest
   

2. Run with llama-server:

   docker run --gpus all -v $(pwd)/models:/model \
       ghcr.io/sowilow/dgx-spark-llama.cpp-bench:latest \
       llama-server -m /model/gpt-oss-20b-q4_mxfp4.gguf -ngl 99 -c 8192
   

Original Model Information

This is a quantized version of openai/gpt-oss-20b. Please refer to the original model card for details on training, safety, and benchmarks.

Citation

@misc{openai2025gptoss120bgptoss20bmodel,
      title={gpt-oss-120b & gpt-oss-20b Model Card}, 
      author={OpenAI},
      year={2025},
      eprint={2508.10925},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2508.10925}, 
}