File	Quantization	Size	RAM	MedQA (vs Unquantized)	Use Case
`medgemma-4b-iq2_xs.gguf`	IQ2_XS (2-bit) + Medical imatrix	~0.85GB	~1.9GB	43.8% (63% retained)	Ultra-budget phones
`medgemma-4b-q2_k.gguf`	Q2_K (2-bit)	~1.3GB	~2.3GB	34.7% (50% retained)	Budget phones
`medgemma-4b-q4_k_m.gguf`	Q4_K_M (4-bit)	~2.3GB	~3GB	56.0% (81% retained)	Budget phones

Medical Importance Matrix

The IQ2_XS model was quantized using a custom importance matrix (imatrix) calibrated on:

African primary care scenarios (malaria, typhoid, cholera, respiratory infections)
Maternal and child health (pregnancy complications, childhood diarrhea, nutrition)
Emergency triage (snake bites, severe dehydration, trauma)
Multi-language symptoms (Twi, Hausa, Yoruba, English)

This preserves medical diagnostic accuracy while aggressively compressing general knowledge.

Usage with llama.cpp

./llama-cli -m medgemma-4b-iq2_xs.gguf -p "Patient has fever, chills, and headache for 3 days. What could this be?"

Built for the Google MedGemma Impact Challenge - bringing AI-powered healthcare to underserved African communities.