BiomedLM
Stanford CRFM 2.7B Biomedical Language Model
Powered by Dr7.ai Medical AI Platform
BioMedLM (formerly PubMedGPT) is a 2.7-billion-parameter GPT-style language model trained by Stanford CRFM and MosaicML exclusively on PubMed abstracts and full-text articles. It shows that a small, domain-specific model can rival much larger general models on biomedical question answering.
Try BioMedLM-Style Biomedical Q&A
Ask questions about biomedical literature, mechanisms or terminology — the demo shows the kind of PubMed-grounded answers a domain-trained language model produces.
Try Other Models
View AllBiomedLM Demo
Experience Advanced Medical AI
Unlock Full BiomedLM Potential
Get unlimited biomedical Q&A, longer context and API access for literature-heavy workflows.
What is BioMedLM?
BiomedLM is a 2.7-billion-parameter autoregressive language model built by the Stanford Center for Research on Foundation Models (CRFM) together with MosaicML. It was first released in December 2022 as PubMedGPT and later renamed BioMedLM.
Unlike general models that see a little biomedical text among everything else, BioMedLM was trained from scratch only on PubMed abstracts and PubMed Central full-text articles (the PubMed portion of The Pile, roughly 50 billion tokens). That focus lets a comparatively tiny model reach about 50% on MedQA (USMLE) and 74% on PubMedQA after fine-tuning — results that at release were competitive with models 50× its size.
BioMedLM is open weights and small enough to fine-tune on a single GPU, which made it a popular base for biomedical extraction, summarisation and question-answering research. It is a research artefact, not a clinical product, and is not intended for patient-facing use.
Latest Development
Released December 2022 as PubMedGPT by Stanford CRFM and MosaicML; renamed BioMedLM in 2023 (paper: Bolton et al., 2024)
Open weights on Hugging Face (stanford-crfm/BioMedLM); trained on 300B tokens of PubMed text
Key Features
A small model trained only on biomedical literature
Core Capabilities
2.7B parameters — fine-tunable on one GPU
Trained exclusively on PubMed abstracts and PMC full text
Biomedical question answering (MedQA, PubMedQA, MedMCQA)
Literature summarisation and information extraction
Biomedical entity and relation extraction after fine-tuning
GPT-2-style architecture with a custom biomedical tokenizer
Open weights for research use
Strong per-parameter efficiency versus general LLMs
Performance Benchmarks
Results reported by Stanford CRFM / MosaicML for the fine-tuned 2.7B model
MedQA (USMLE)
50.3%US Medical Licensing Examination-style questions, 4-option
PubMedQA
74.4%Yes/no/maybe questions over PubMed abstracts
MedMCQA
CompetitiveIndian medical entrance exam questions; reported in the BioMedLM paper
Model size
2.7BReaches these scores with a fraction of the parameters of general LLMs
Use Cases
Biomedical Research
Summarise papers, extract entities and generate hypotheses from the literature.
Literature Analysis
Question answering and summarisation grounded in PubMed-style text.
Medical Education
Answer exam-style biomedical questions and explain underlying mechanisms.
How to Use BioMedLM
Open weights that fit on a single GPU
Get the Weights
Load stanford-crfm/BioMedLM from Hugging Face with transformers. The base model is a plain language model — fine-tune it for your task.
Integration Options
BioMedLM is most useful as a base for task-specific fine-tuning.
- • Fine-tune for biomedical QA or classification
- • Named-entity and relation extraction
- • Abstract summarisation pipelines
- • Domain-adapted embeddings for retrieval
Deployment Considerations
Plan for privacy, validation and human oversight before putting BioMedLM anywhere near patient data.
- • De-identify data and comply with HIPAA / GDPR
- • Validate on your own institution's data before use
- • Keep a licensed clinician in the loop for every output
- • Document limitations and intended use for auditors
Important Considerations
Clinical Validation Required
BioMedLM is a research model. Its outputs must be reviewed by qualified healthcare professionals before any clinical use.
Regulatory Compliance
The model is not a cleared or approved medical device. Ensure compliance with local healthcare regulations and data-protection rules (e.g. HIPAA, GDPR) before deployment.
BioMedLM vs Other Medical Language Models
Small and specialised versus large and general
BioMedLM
Stanford CRFM's 2.7B model
- ✓Trained only on biomedical literature
- ✓Runs and fine-tunes on a single GPU
- ✓Open weights, fully inspectable
- ✓Competitive MedQA / PubMedQA scores for its size
- ✓Custom biomedical tokenizer
- ✓Well-documented training recipe
Research groups that need a fine-tunable biomedical base model on modest hardware
Other Medical AI Models
How it compares
- ×General LLMs (GPT-4, Gemini) are stronger at open-ended reasoning but are closed and costly
- ×Larger open models (Meditron-70B, MedGemma-27B) need far more GPU memory
- ×Text-only models cannot look at images
- ×Image-only classifiers cannot explain their findings in language
- ×Commercial products are validated but not inspectable
- ×Most alternatives publish fewer reproducible benchmarks
Pick by modality (text vs. image), licence, and the hardware you actually have
Frequently Asked Questions
Common questions about BioMedLM
Who developed BioMedLM?
Stanford CRFM (Center for Research on Foundation Models) together with MosaicML. It was originally released as PubMedGPT in December 2022.
How big is BioMedLM?
2.7 billion parameters — small enough to fine-tune on one GPU, which is the whole point of the project.
What did BioMedLM score on medical exams?
About 50.3% on MedQA (USMLE) and 74.4% on PubMedQA after fine-tuning, according to the Stanford CRFM release.
Can I use BioMedLM in a clinical product?
The weights are open for research. It is not a medical device and the authors explicitly caution against patient-facing use without validation.
Does this demo run BioMedLM itself?
The demo uses Dr7.ai's medical assistant to show PubMed-style biomedical Q&A. For exact reproductions, run the open weights locally.