이 모델의 강점은 무엇인가요?

의료 분야 특화 최적화 넓은 128K 컨텍스트 윈도우 DeepMind의 높은 신뢰도

이 모델의 약점은 무엇인가요?

비공개 라이선스 시스템 높은 연산 자원 요구사항 범용 작업 적응성

어떤 용도에 가장 적합한가요?

의료 정보 고급 분석 전문 의료 대화 대규모 의료 문헌 처리

모델 목록으로

DeepMind독점

MedGemma 27B Instruct

Name: MedGemma 27B Instruct
Author: DeepMind

MedGemma 27B Instruct은 DeepMind가 개발한 의료 분야 특화 기반 모델입니다. 270억 파라미터 규모이며, 최대 128K의 긴 컨텍스트 윈도우를 지원하는 대화형 모델입니다.

파라미터

270.0B

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

・의료 분야 특화 최적화
・넓은 128K 컨텍스트 윈도우
・DeepMind의 높은 신뢰도

약점

・비공개 라이선스 시스템
・높은 연산 자원 요구사항
・범용 작업 적응성

활용 사례

・의료 정보 고급 분석
・전문 의료 대화
・대규모 의료 문헌 처리

심층 분석

Parameters

27B

Available in text-only and multimodal variants

MedQA

87.7%

Within 3 points of DeepSeek R1 at ~1/10th the inference cost

Chest X-ray Reports

81% clinically sufficient

US board-certified radiologist review for 4B variant

Base Model

Gemma 3

Medically optimized variant of Gemma 3

Release Date

July 9, 2025

Part of Health AI Developer Foundations (HAI-DEF)

License

HAI-DEF Terms of Use

Special health AI license, not standard Gemma

Deployment

Single GPU

Can run on a single GPU; 4B variant runs on mobile

Modalities

Text + Images

Multimodal: medical text and imaging (X-rays, dermatology, ophthalmology, histopathology)

강점

・Best-in-class MedQA (87.7%) among models under 50B parameters
・Within 3 points of DeepSeek R1 at 1/10th the inference cost
・Multimodal: handles medical images (X-rays, dermatology, ophthalmology, histopathology)
・Retains general Gemma capabilities (instruction-following, non-English languages)
・Specialized SigLIP image encoder pre-trained on medical data
・Runs on a single GPU

약점

・Not a general-purpose model - specialized for medical tasks
・HAI-DEF license more restrictive than Apache 2.0
・Requires fine-tuning for optimal performance on specific medical tasks
・Limited to text and image output (no audio/video)
・May not be suitable for direct clinical use without validation

경쟁사 비교

Model	Arena	SWE	GPQA	Price
MedGemma 27B	N/A	N/A	87.7% (MedQA)	Free (open weights)
DeepSeek R1	~1470	~70%	~90% (MedQA)	Free (open weights)
BioMistral 7B	N/A	N/A	~60% (MedQA)	Free (open weights)
GPT-5 (general)	~1480	~80%	~85% (MedQA est)	$5/$20 per 1M

개요

MedGemma 27B is Google's most capable open medical model, part of the Health AI Developer Foundations (HAI-DEF) collection. Available in text-only and multimodal variants, it achieves 87.7% on MedQA - within 3 points of DeepSeek R1 at approximately 1/10th the inference cost. Specialized for medical text comprehension, report generation, and medical image interpretation.

벤치마크 및 성능

MedQA: 87.7% (among best under 50B). Competitive with larger models on medical retrieval and EHR interpretation. The 4B variant (separate model) achieved 64.4% MedQA and 81% clinically sufficient chest X-ray reports. After fine-tuning, achieves RadGraph F1 of 30.3 on chest X-ray report generation.

상세 비교

Near DeepSeek R1 medical performance at dramatically lower cost. Much smaller than general-purpose frontier models but superior on medical tasks within its size class. The 27B text variant is the strongest; multimodal variant adds image understanding.

커뮤니티 평가

Well-received by healthcare AI developers. The open nature enables privacy-preserving deployment. Used for medical report generation, clinical decision support, and medical image analysis. Researchers appreciate the specialized SigLIP encoder for medical imaging.

활용 사례

Medical text reasoning, clinical report generation, medical image interpretation (X-rays, dermatology, ophthalmology, histopathology), EHR data analysis, medical question answering. Designed as a starting point for healthcare AI development with fine-tuning for specific clinical tasks.