Use case 7 of 13

A medical literature helper (not medical advice)

Search and summarise PubMed papers with a medical model, privately, as a reading aid for clinicians, researchers and curious patients.

Stacked books, reading glasses and a blue stethoscope beside a closed laptop — Image: generated with Grok Imagine

The real example

PubMed + MedGemma, with Google's own guard-rails

PubMed holds more than 40 million citations for biomedical literature, and NCBI offers documented programmatic access through its E-utilities APIs. Google's MedGemma models are tuned for medical text and images; on MedQA (4-option) the model card lists 64.4% for MedGemma 1 4B, 69.1% for MedGemma 1.5 4B and 85.3% for MedGemma 1 27B (multimodal). The same card is explicit that outputs are not intended to directly inform clinical diagnosis, patient management decisions or treatment recommendations, and must be independently verified.

Sources: PubMed · NCBI APIs (E-utilities) · MedGemma model card

Why run it locally

Questions about your own or a patient's condition are sensitive health data. A local tool keeps the questions private and works on a hospital network without outside access.

Recommended stack

Three tiers built from our builds. Weights-in-memory is our arithmetic (parameters × bits), checked against each device's usable memory; it excludes KV cache and runtime overhead. Runtime compatibility is not validated.

Budget

Value 16 GB

$1,173Sample

16 GB on GeForce RTX 5060 Ti 16 GB

MedGemma 1.5 4B · 2.6 GB (~4-bit)
Qwen3 Embedding 0.6B · 1.2 GB (16-bit)

Weights ≈ 3.8 GB of 15.5 GB usableEstimate

Excludes KV cache and runtime overhead. Runtime compatibility is not validated.

The 4B model plus an embedder fit easily; good for summarising abstracts.

Runtime: Ollama or llama.cpp + a small RAG script over PubMed abstracts

Mid

Used-3090 24 GB

$1,980Sample

24 GB on GeForce RTX 3090 24 GB (used)

MedGemma 27B text · 16 GB (~4-bit)
Qwen3 Embedding 0.6B · 1.2 GB (16-bit)

Weights ≈ 17.6 GB of 23.5 GB usableEstimate

Excludes KV cache and runtime overhead. Runtime compatibility is not validated.

The 27B text model at 4-bit is tight on 24 GB; keep context modest.

Runtime: llama.cpp server

Pro

Dual-GPU 48 GB

$4,044Sample

48 GB on 2× RTX 3090 (used) 48 GB

MedGemma 27B · 17 GB (~4-bit)
Qwen3 Embedding 8B · 4.6 GB (~4-bit)
Qwen3 Reranker 0.6B · 1.2 GB (16-bit)

Weights ≈ 23.3 GB of 47.0 GB usableEstimate

Excludes KV cache and runtime overhead. Runtime compatibility is not validated.

Multimodal 27B with room for long papers and figures.

Runtime: vLLM + RAG pipeline

Setup outline

Accept the Health AI Developer Foundations terms on Hugging Face to download MedGemma.
Pull abstracts for your topic via E-utilities (respect NCBI's usage limits) and embed them locally.
Answer questions only from retrieved abstracts, always showing PMIDs and links.
Add a fixed banner and a system prompt stating the tool is a reading aid, not medical advice.

Limits

Not medical advice and not a medical device. Google's own model card rules out direct clinical use.
Abstracts are not full papers. Many key details sit behind paywalls.
MedQA measures exam-style questions, not real clinical safety.

Next: A patient maths tutor

Build prices come from our build datasheets: GPUs and devices are sourced retail or sold prices; other parts are sample prices until sourced, and the flag on each total says which. Model links go to our datasheets, where each parameter count and licence was checked against the Hugging Face API.