Instructions to use bbarn4/medgemma-27b-text-it-GPTQ with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use bbarn4/medgemma-27b-text-it-GPTQ with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="bbarn4/medgemma-27b-text-it-GPTQ")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("bbarn4/medgemma-27b-text-it-GPTQ")
model = AutoModelForCausalLM.from_pretrained("bbarn4/medgemma-27b-text-it-GPTQ")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use bbarn4/medgemma-27b-text-it-GPTQ with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "bbarn4/medgemma-27b-text-it-GPTQ"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "bbarn4/medgemma-27b-text-it-GPTQ",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/bbarn4/medgemma-27b-text-it-GPTQ

SGLang

How to use bbarn4/medgemma-27b-text-it-GPTQ with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "bbarn4/medgemma-27b-text-it-GPTQ" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "bbarn4/medgemma-27b-text-it-GPTQ",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "bbarn4/medgemma-27b-text-it-GPTQ" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "bbarn4/medgemma-27b-text-it-GPTQ",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use bbarn4/medgemma-27b-text-it-GPTQ with Docker Model Runner:
```
docker model run hf.co/bbarn4/medgemma-27b-text-it-GPTQ
```

MedGemma-27B-Text-IT GPTQ (4-bit)

This is a 4-bit GPTQ quantization of google/medgemma-27b-text-it, the text-only variant of Google's MedGemma optimized for medical text reasoning.

Created For

This quantized model was generated by Ben Barnard Ph.D and Oladimeji Adaramewa for MPART — the Medical Policy Applied Research Team at Innovate Springfield and University of Illinois Springfield. MPART focuses on applied research in healthcare policy, Medicaid concerns, and health system funding analysis.

Why this model?

The text-only MedGemma 27B scores higher on medical text benchmarks than the multimodal version (89.8 vs 87.0 on MedQA, 74.2 vs 70.2 on MedMCQA) while being simpler to deploy. This quantized version reduces memory requirements from ~55GB to ~15GB, making it runnable on a single GPU with 24GB+ VRAM. We found that it is very good with understanding healthcare policy and finance.

Quantization Details

Method: GPTQ via llmcompressor (Neural Magic)
Bits: 4 (W4A16 — 4-bit weights, 16-bit activations)
Group size: 128
Calibration data: 256 samples from C4 (English)
Ignored layers: lm_head (kept at full precision)

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model = AutoModelForCausalLM.from_pretrained(
    "bbarn4/medgemma-27b-text-it-GPTQ",
    device_map="auto",
    dtype="auto"
)
tokenizer = AutoTokenizer.from_pretrained("bbarn4/medgemma-27b-text-it-GPTQ")

prompt = "What are the key differences between Type 1 and Type 2 diabetes?"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.inference_mode():
    output = model.generate(**inputs, max_new_tokens=500, do_sample=False)

print(tokenizer.decode(output[0], skip_special_tokens=True))

Base Model Performance (pre-quantization)

These are the published scores for the full-precision base model:

Benchmark	MedGemma 27B Text	Gemma 3 27B
MedQA (4-op)	89.8	74.9
MedMCQA	74.2	62.6
PubMedQA	76.8	73.4
MMLU Med	87.0	83.3
MedXpertQA	25.7	15.7

Intended Use

This model is intended as a starting point for developers and researchers building healthcare applications involving medical text. It is NOT intended for direct clinical use. All outputs require independent verification by qualified professionals.

License

Use is governed by the Health AI Developer Foundations terms of use.

Downloads last month: 3,132

Safetensors

Model size

5B params

Tensor type

I64

I32

BF16

Model tree for bbarn4/medgemma-27b-text-it-GPTQ

Base model

google/gemma-3-27b-pt

Finetuned

google/medgemma-27b-text-it

Quantized

(26)

this model