MedGemma 27B
Medical-specialist Gemma fine-tune. Trained on de-identified medical literature and imaging. Research use under HAI-DEF terms.
Our verdict
Positioning
MedGemma 27B is Google's medical-domain fine-tune of Gemma — designed for clinical text understanding, medical QA, and healthcare NLP tasks. Built on the Gemma 4 31B architecture with medical-domain pretraining and instruction tuning, it targets healthcare deployments that need an on-prem model for privacy-sensitive clinical data. The operator-grade question: "does medical specialization at 27B deliver enough domain accuracy to be useful in clinical settings, or is a general-purpose 31B model with good prompting close enough?"
Strengths
- Medical-domain specialization. Trained on clinical texts, medical literature, and healthcare QA datasets. Medical terminology, drug names, procedure codes, and clinical reasoning are meaningfully better than general-purpose models at this size.
- Privacy-preserving on-prem deployment. Healthcare data often cannot leave the institution's servers. MedGemma 27B runs on a single 24-GB card — deployable within hospital firewalls.
- Fits a single 24-GB consumer card at Q4. Q4_K_M at ~17 GB — comfortable on RTX 4090, RTX 3090. Leaves headroom for long clinical documents.
- Google's medical AI research backing. MedGemma benefits from Google's Med-PaLM research lineage and medical-domain evaluation rigor. Clinical benchmark scores are validated against medical professional standards.
- Good for specific medical tasks. Medical summarization, clinical note generation, patient education material drafting, ICD/CPT code suggestions — structured medical tasks where domain vocabulary matters.
- Permissive Gemma license. Less restrictive than Llama for healthcare deployments. Verify terms for clinical use.
Limitations
- Not a medical device. MedGemma 27B is not FDA-approved, not clinically validated for diagnosis, and should not be used for autonomous clinical decision-making. This is an assistive tool, not a replacement for medical judgment.
- Medical knowledge is general, not specialty-deep. While stronger than general-purpose models, MedGemma's clinical depth in subspecialties (neurosurgery, interventional cardiology, etc.) is limited. Specialist physician review is mandatory.
- Hallucinations in clinical contexts are dangerous. All LLMs hallucinate. In medical contexts, a hallucinated drug dosage or fabricated study reference can cause harm. MedGemma reduces but does not eliminate this risk.
- Knowledge cutoff is training-dependent. Medical knowledge evolves rapidly. New drugs, guidelines, and studies post-cutoff are absent. Living medical QA systems need RAG on current literature.
- English-only medical training. Non-English clinical terminology and medical guidelines are not covered. For multilingual healthcare deployments, consider the general-purpose Qwen 3 32B with medical prompting.
- Regulatory compliance is your responsibility. HIPAA, GDPR, and regional healthcare data regulations apply to your deployment. MedGemma is a model — compliance is an infrastructure concern.
- Smaller community than general-purpose Gemma. Fewer fine-tunes, fewer deployment guides specific to healthcare workflows. Self-serve support burden is higher.
Real-world performance
- vs Gemma 4 31B: MedGemma 27B wins on medical tasks by meaningful margins — better clinical terminology, better differential diagnosis reasoning, better drug interaction knowledge. For non-medical tasks, Gemma 4 31B is the better generalist.
- vs Qwen 3 32B: Qwen 3 32B with careful medical prompting can approach MedGemma's quality on some tasks, but MedGemma's medical pretraining gives it an edge on clinical terminology and structured medical reasoning.
- vs Llama 3.3 70B: Llama 3.3 70B at 70B has more raw reasoning power but no medical specialization. For clinical tasks where domain knowledge matters more than general reasoning, MedGemma often wins despite the parameter disadvantage.
- vs OLMo 2 32B: OLMo 2 is general-purpose, fully open. MedGemma is medical-specialized, closed training. Different domains, different openness.
Should you run this locally?
Yes if you're deploying in a healthcare setting with privacy constraints, need on-prem medical NLP, and have clinical oversight for all model outputs. MedGemma 27B is the best open-weight medical model at the single-GPU tier.
No if you're building a patient-facing diagnosis tool without clinical oversight. This is not safe. Medical LLMs are assistive, not autonomous.
Probably not if you need general medical knowledge plus broad language support. Qwen 3 32B with RAG on medical literature may be more practical for multilingual healthcare deployments.
How it compares
- vs Gemma 4 31B: MedGemma is the medical specialist; Gemma 4 31B is the generalist. Same family, different fine-tuning targets.
- vs Qwen 3 32B: Qwen is general-purpose with broader languages. MedGemma is medical-specific with deeper domain knowledge. Pick by domain.
- vs OLMo 2 32B: OLMo 2 is transparent general-purpose. MedGemma is specialized medical. Different optimization targets.
- vs Llama 3.3 70B: Llama has more capability at higher VRAM. MedGemma has medical domain advantage at lower VRAM.
Run this yourself
Overview
Medical-specialist Gemma fine-tune. Trained on de-identified medical literature and imaging. Research use under HAI-DEF terms.
How to run it
MedGemma 27B is Google's medical-domain fine-tune of Gemma 27B. Specialized for clinical text understanding, medical Q&A, and biomedical reasoning. Run at Q4_K_M via Ollama (ollama pull medgemma:27b) or llama.cpp with -ngl 999 -fa -c 4096. Q4_K_M file size ~15 GB on disk. Minimum VRAM: 12 GB — RTX 4070 (12GB) at Q4_K_M with KV offload for 4K context. RTX 4090 24GB: Q4_K_M comfortably at 16K+ context. Recommended: RTX 4090 24GB at Q4_K_M. Throughput: ~40-60 tok/s on RTX 4090 at Q4_K_M. Gemma architecture — well-supported in llama.cpp, Ollama, and Gemma-specific stacks. MedGemma is Google's medical AI offering — trained on medical literature, clinical notes, and biomedical datasets. Strong on: medical terminology, diagnostic reasoning, drug information, clinical summarization. Weak on: non-medical topics (catastrophic forgetting from domain fine-tuning), current medical guidelines (knowledge cutoff). Not FDA-approved — research model only. License: Gemma license (verify commercial terms). Use for: medical research, clinical decision support (with human review), biomedical literature analysis. Never for: autonomous clinical decisions. Context: Gemma's 8K (practical 4-8K on 12-16 GB). Medical contexts are typically shorter — less KV pressure.
Hardware guidance
Minimum: RTX 3060 12GB at Q3_K_M with KV offload. Recommended: RTX 4090 24GB at Q4_K_M (16K+ context). VRAM math: 27B dense, Q4_K_M ≈ 15 GB. KV cache at 8K: ~6 GB. Total: ~21 GB at 8K. RTX 4090 24GB: Q4 + 8-16K context — comfortable on-GPU. RTX 3080 10GB: Q3_K_M with KV offload. RTX 4080 16GB: Q4 + 4K context on-GPU. MacBook Pro M4 Pro 24GB+: Q4 at 12-25 tok/s. Cloud: A10 24GB at Q4_K_M. MedGemma is one of the largest medical-specific models that fits comfortably on consumer GPUs. AWQ-INT4 drops to ~13 GB. For clinical settings: the model's knowledge is frozen — pair with RAG on current medical literature (PubMed, UpToDate).
What breaks first
- Medical hallucination. MedGemma generates plausible-sounding but incorrect medical information — drug dosages, contraindications, treatment protocols. Never use for autonomous clinical decisions. Always verify outputs against current medical guidelines. 2. Catastrophic forgetting. Domain fine-tuning on medical data severely degrades general knowledge. The model will hallucinate on non-medical topics worse than base Gemma 27B. 3. Knowledge cutoff. Medical knowledge is frozen at training time. New drugs, treatment guidelines, and research published after the cutoff aren't known. RAG on current literature is essential for clinical use. 4. Q3 terminology precision. Medical terminology at Q3 introduces dangerous errors — confusing similar drug names, dosages, or anatomical references. Use Q4_K_M minimum for medical tasks. Q8 or FP16 for clinical settings.
Runtime recommendation
Common beginner mistakes
Mistake: Using MedGemma as a diagnostic tool without physician review. Fix: This is a research model, not a medical device. Always have a qualified physician review outputs. The model hallucinates drug names and dosages. Mistake: Trusting MedGemma for current treatment guidelines. Fix: Knowledge cutoff limits the model to pre-training data. Use RAG on UpToDate, PubMed, or current clinical guidelines for up-to-date information. Mistake: Using Q3 quantization for medical tasks. Fix: Q3 confuses medical terminology. Use Q4_K_M minimum. Q8 or FP16 for clinical settings where precision matters. Mistake: Expecting MedGemma to perform well on general knowledge tasks. Fix: Medical fine-tuning causes catastrophic forgetting of non-medical knowledge. Use base Gemma 27B for non-medical tasks.
Family & lineage
How this model relates to others in its lineage. Family members share architecture and training-data roots; parent / children edges record direct distillation or fine-tune relationships.
Strengths
- Medical-domain accuracy
Weaknesses
- Not for clinical decisions
- License restrictions
Quantization variants
Each quantization trades model quality for file size and VRAM. Q4_K_M is the most popular starting point.
| Quantization | File size | VRAM required |
|---|---|---|
| Q4_K_M | 16.0 GB | 20 GB |
Get the model
HuggingFace
Original weights
Source repository — direct quantization required.
Hardware that runs this
Cards with enough VRAM for at least one quantization of MedGemma 27B.
Models worth comparing
Same parameter band, plus what's one tier above and below — so you can decide what actually fits your hardware.
Frequently asked
What's the minimum VRAM to run MedGemma 27B?
Can I use MedGemma 27B commercially?
What's the context length of MedGemma 27B?
Does MedGemma 27B support images?
Source: huggingface.co/google/medgemma-27b-it
Reviewed by RunLocalAI Editorial. See our editorial policy for how we research and verify model claims.
Related — keep moving
Verify MedGemma 27B runs on your specific hardware before committing money.