이 모델의 강점은 무엇인가요?

충분한 4B 파라미터 규모 광대한 128K 컨텍스트 처리 의료 전문 채팅 최적화

이 모델의 약점은 무엇인가요?

비공개 라이선스로 인한 제한 규모로 인한 높은 리소스 소모 범용 작업에 대한 불명확한 적응성

어떤 용도에 가장 적합한가요?

의료 정보 챗봇 개발 긴 의료 논문 분석/요약 전문 의료 대화 생성

모델 목록으로

DeepMind독점

MedGemma 4B Instruct

Name: MedGemma 4B Instruct
Author: DeepMind

MedMind이 개발한 MedGemma 4B Instruct는 의료 분야에 특화된 기반 모델이며, 채팅 형식으로 제공됩니다. 약 400억 파라미터 규모를 가지며, 최대 128K의 긴 컨텍스트 창을 지원합니다.

파라미터

40.0B

컨텍스트

128K

라이선스

https://creativecommons.org/licenses/by-sa/4.0/

출시일

2025-07-09

API 가격

이 모델의 API 가격 정보는 현재 공개되지 않았습니다

강점

・충분한 4B 파라미터 규모
・광대한 128K 컨텍스트 처리
・의료 전문 채팅 최적화

약점

・비공개 라이선스로 인한 제한
・규모로 인한 높은 리소스 소모
・범용 작업에 대한 불명확한 적응성

활용 사례

・의료 정보 챗봇 개발
・긴 의료 논문 분석/요약
・전문 의료 대화 생성

심층 분석

Parameters

Multimodal variant, can run on mobile hardware

MedQA

64.4%

Among best very small (<8B) open models

Chest X-ray Accuracy

81% clinically sufficient

US board-certified radiologist judged reports comparable to original

Base Model

Gemma 3

Medically optimized variant with SigLIP image encoder

Release Date

July 9, 2025

Part of HAI-DEF collection

Deployment

Mobile-capable

Can be adapted to run on mobile hardware

Image Encoder

MedSigLIP

Pre-trained on chest X-rays, dermatology, ophthalmology, histopathology

License

HAI-DEF Terms of Use

Health AI Developer Foundations license

강점

・Runs on mobile hardware - extremely accessible medical AI
・64.4% MedQA is excellent for a 4B model
・81% clinically sufficient chest X-ray reports after radiologist review
・Multimodal: handles both medical text and images
・Retains general Gemma capabilities for mixed medical/non-medical tasks
・Competitive with task-specific state-of-the-art on medical image classification

약점

・Significantly lower MedQA than 27B variant (64.4% vs 87.7%)
・HAI-DEF license more restrictive than Apache 2.0
・Requires fine-tuning for best performance on specific tasks
・Limited to text and image modalities
・Not suitable for complex multi-step medical reasoning

경쟁사 비교

Model	Arena	SWE	GPQA	Price
MedGemma 4B	N/A	N/A	64.4% (MedQA)	Free (open weights)
MedGemma 27B	N/A	N/A	87.7% (MedQA)	Free (open weights)
BioMistral 7B	N/A	N/A	~60% (MedQA)	Free (open weights)
Gemma 3 4B (base)	~1200	N/A	~45% (MedQA est)	Free (open weights)

개요

MedGemma 4B is Google's compact medical AI model that can run on mobile hardware. With 64.4% MedQA (among best under 8B) and 81% clinically sufficient chest X-ray reports, it provides accessible medical AI for resource-constrained deployments. Part of the Health AI Developer Foundations collection.

벤치마크 및 성능

MedQA 64.4% (best-in-class for <8B models). 81% of generated chest X-ray reports judged clinically sufficient by US board-certified radiologist. Competitive with task-specific SOTA on medical image classification. After fine-tuning, achieves RadGraph F1 of 30.3 on chest X-ray generation.

상세 비교

Much more accessible than MedGemma 27B (runs on mobile vs single GPU). Lower raw medical knowledge (64.4% vs 87.7% MedQA) but still excellent for its size. Better than base Gemma 3 4B on medical tasks by a wide margin.

커뮤니티 평가

Popular for mobile health apps and point-of-care AI tools. The chest X-ray report generation capability is frequently cited. Developers appreciate the ability to run medical AI locally without cloud dependencies.

활용 사례

Mobile medical AI applications, point-of-care decision support, chest X-ray report generation, medical image classification, dermatology screening, and privacy-preserving local medical AI. Ideal for clinics and hospitals that need on-device processing.