license: apache-2.0 language: - en tags: - gguf - quantized - gpt-oss - llama-cpp - mixture-of-experts - f16 model_type: gpt-oss

GPT-OSS 20B F16 GGUF

This is a high-quality F16 GGUF conversion of the GPT-OSS 20B model, optimized for llama.cpp inference.

Model Details

Model Type: GPT-OSS (Mixture of Experts)
Parameters: 20.91B total, 1.8B active
Precision: F16 (16-bit floating point)
File Size: 12.83 GiB
Context Length: 131,072 tokens
Experts: 32 total, 4 active per token
Architecture: Sliding window attention with expert routing

Key Features

✅ Fully Functional: All 459 tensors intact, including mxfp4 expert weights
✅ High Quality: F16 precision maintains model performance
✅ Complete MoE Support: Expert routing and gating fully preserved
✅ Extended Context: 131K token context window with YARN scaling

Usage with llama.cpp

# Download the model
huggingface-cli download davidfred/gpt-oss-20b-f16-gguf gpt-oss-20B-F16.gguf


## Hardware Requirements

- **Minimum RAM**: 16 GB (recommended: 24 GB+)
- **CPU**: Multi-core recommended (tested on 8 vCPU)
- **Storage**: ~13 GB free space
- **OS**: Compatible with llama.cpp (Linux, Windows, macOS)

## Performance Notes

- Efficiently runs on CPU-only setups
- Utilizes mixture of experts for optimal parameter efficiency
- Supports both interactive and batch inference modes
- Compatible with llama.cpp server mode for API access

## Model Origin

Converted from the original GPT-OSS 20B model using llama.cpp conversion tools. This F16 GGUF preserves all model capabilities while providing efficient inference performance.

## License

Apache 2.0 - Same as the original GPT-OSS model.
EOF

# Upload the README
huggingface-cli upload davidfred/gpt-oss-20b-f16-gguf /tmp/README.md README.md

./llama-cli -m gpt-oss-20B-F16.gguf --prompt "Your prompt here" -n 128 --threads 8



## Hardware Requirements

- **Minimum RAM**: 16 GB (recommended: 24 GB+)
- **CPU**: Multi-core recommended (tested on 8 vCPU)
- **Storage**: ~13 GB free space
- **OS**: Compatible with llama.cpp (Linux, Windows, macOS)

## Performance Notes

- Efficiently runs on CPU-only setups
- Utilizes mixture of experts for optimal parameter efficiency
- Supports both interactive and batch inference modes
- Compatible with llama.cpp server mode for API access

## Model Origin

Converted from the original GPT-OSS 20B model using llama.cpp conversion tools. This F16 GGUF preserves all model capabilities while providing efficient inference performance.

## License

Apache 2.0 - Same as the original GPT-OSS model.
EOF

# Upload the README
huggingface-cli upload davidfred/gpt-oss-20b-f16-gguf /tmp/README.md README.md
Quick Upload Script (Alternative)
If you prefer a single script approach:

cat > /tmp/upload_model.py << 'EOF'
from huggingface_hub import HfApi, create_repo
import os

# Configuration
repo_id = "davidfred/gpt-oss-20b-f16-gguf"
model_path = os.path.expanduser("~/openai/gpt-oss-20b/gpt-oss-20B-F16.gguf")

# Create repository
create_repo(repo_id, exist_ok=True)

# Initialize API
api = HfApi()

# Upload main model file
print("Uploading main model file...")
api.upload_file(
    path_or_fileobj=model_path,
    path_in_repo="gpt-oss-20B-F16.gguf",
    repo_id=repo_id,
    commit_message="Add GPT-OSS 20B F16 GGUF model"
)

print(f"✅ Model uploaded successfully!")
print(f"🔗 Repository: https://huggingface.co/{repo_id}")
print(f"📁 Direct download: https://huggingface.co/{repo_id}/resolve/main/gpt-oss-20B-F16.gguf")
EOF

python /tmp/upload_model.py