Get started with Standard Model Biomedicine’s biological world model. This guide walks you through setting up your environment and downloading SMB-v1-Structure.
New: This quickstart now features SMB-v1-Structure — our flagship JEPA-based multimodal foundation model for oncology.Prerequisites GPU support is strongly recommended. SMB-v1-Structure (1.7B parameters) requires approximately 16GB GPU memory for inference.
Before you begin, ensure you have the following installed:
Python 3.10+ — Required for running the modelspip — Python package managerCUDA (recommended) — For GPU acceleration with NVIDIA GPUsGit — For cloning repositories
One-Command Setup (Recommended) Run the quickstart script to automatically configure your environment:
curl -fsSL https://raw.githubusercontent.com/standardmodelbio/quickstart/main/quickstart.sh -o quickstart.sh && source quickstart.sh
This script will:
Create a Python 3.10 virtual environment named standard_model
Install PyTorch with CUDA support (if available)
Install HuggingFace libraries (transformers, datasets, accelerate)
Install SMB utilities (smb-biopan-utils)
Download the Standard Model to your local machine
Manual Installation If you prefer to set up your environment manually:
Create Virtual Environment
python3 -m venv standard_model source standard_model/bin/activate
Install PyTorch
CUDA 12.x
CUDA 11.x
CPU Only
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
Install Dependencies
pip install transformers datasets accelerate huggingface_hub pandas pip install git+https://github.com/standardmodelbio/smb-biopan-utils.git
Download SMB-v1-Structure
from huggingface_hub import snapshot_download snapshot_download( "standardmodelbio/SMB-v1-1.7B-Structure" )
Environment Activation After setup, activate your environment for usage:
source standard_model/bin/activate
Verify Installation Verify that everything is working correctly:
import torch from transformers import AutoModelForCausalLM, AutoTokenizer # Check PyTorch and CUDA print ( f "PyTorch version: { torch. __version__ } " ) print ( f "CUDA available: { torch.cuda.is_available() } " ) # Load SMB-v1-Structure model_id = "standardmodelbio/SMB-v1-1.7B-Structure" tokenizer = AutoTokenizer.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained( model_id, trust_remote_code = True , device_map = "auto" ) print ( "SMB-v1-Structure loaded successfully!" ) print ( f "Model parameters: { sum (p.numel() for p in model.parameters()) :,} " )
SMB-v1-Structure is a world model , not a text generator. It predicts patient states in latent space, not tokens. See the Embeddings Guide for the full workflow. Download Other Models Download additional models from the Standard Model family:
SMB-EHR-4B
SMB-Vision-Base
SMB-Language-8B
from huggingface_hub import snapshot_download snapshot_download( "standardmodelbio/smb-ehr-4b" )
Troubleshooting
CUDA Not Detected Ensure NVIDIA drivers are up to date. Run nvidia-smi to verify GPU is accessible.
Out of Memory SMB-v1-Structure requires ~16GB GPU memory. Use torch.float16 or quantization for smaller GPUs.
Model Access Denied Some models may require authentication. Run huggingface-cli login with your token.
Slow Download Model downloads can be large (several GB). Ensure stable connection and sufficient disk space.
Reducing Memory Usage For GPUs with less memory, use half-precision or quantization:
import torch from transformers import AutoModelForCausalLM # Load in bfloat16 (half precision + f32 range) model = AutoModelForCausalLM.from_pretrained( "standardmodelbio/SMB-v1-1.7B-Structure" , trust_remote_code = True , torch_dtype = torch.bfloat16, device_map = "auto" ) # Or use 8-bit quantization (requires bitsandbytes) # pip install bitsandbytes model = AutoModelForCausalLM.from_pretrained( "standardmodelbio/SMB-v1-1.7B-Structure" , trust_remote_code = True , load_in_8bit = True , device_map = "auto" )
Hardware Requirements Model Parameters Min GPU Memory Recommended SMB-v1-Structure 1.7B 16 GB 32 GB SMB-EHR-4B 4B 24 GB 48 GB SMB-Vision-Base 97.2M 4 GB 8 GB SMB-Language-8B 8B 32 GB 80 GB
Next Steps 