MedGemma 1.5 GGUF (Optimized for Intel Mac & Metal)

Project Overview

Quantization Details

• Method: Quantized using llama.cpp
• Format: GGUF
• Quantization Level: Q4_K_M (Optimal balance between clinical accuracy and memory efficiency)
• Target Hardware: Intel Mac with 16GB+ RAM and AMD GPU (Metal enabled)

Performance on Edge Hardware

• Inference Speed: ~15 tokens/sec
• Reasoning Latency: Sub-second response for clinical triage logic.
• Memory Footprint: Fits comfortably within 16GB RAM constraints.

Intended Use

• Clinical Triage Assistance: Helping frontline workers prioritize patients based on MSF and US Army protocols.
• Offline Medical RAG: Serving as the reasoning engine for local vector databases (ChromaDB).
• Privacy-First Applications: Scenarios where patient data residency is strictly local.

How to Use

1. Download the .gguf file.
2. Use with llama.cpp or any GGUF-compatible runner.
3. For Metal acceleration on Mac:
   ./llama-cli -m medgemma-1.5-q4_k_m.gguf -ngl 99 --color -p "Patient presents with..."

Disclaimers

IMPORTANT: This model is for research and educational purposes only as part of the MedGemma Impact Challenge. It is not intended for final clinical diagnosis or as a substitute for professional medical advice, diagnosis, or treatment. Always consult with a qualified healthcare provider.

Acknowledgements

• Base Model: Google MedGemma 1.5
• Competition: The MedGemma Impact Challenge (Kaggle)
• Developer: Edmond Song (Stardust Kei)