# Models

Technical documentation for biomedical world models and modality-specific encoders. Extract patient embeddings, fuse modalities, and build clinical prediction models for oncology and healthcare.

| | |
|---|---|
| Latest release | SMB-v1 — 5 variants · 1.7B / 8B |
| License | Open weights, research use |
| Modalities | 5 — EHR · CT · WGS · Pathology · Proteomics |

All weights are published on [Hugging Face](https://huggingface.co/standardmodelbio). There is no hosted service to call.

## smb-series

World-model encoders trained on longitudinal oncology data. Extract patient embeddings that are predictive across nearly every clinical task — readmission, staging, response, survival.

Trained with JEPA-SFT or curriculum training on longitudinal EHR from Memorial Sloan Kettering — 20K+ patients, 300K+ patient-years, 7 cancer types.

| Variant | Backbone | Paradigm |
|---|---|---|
| [SMB-v1_Qwen3-1.7b_multi-objective](https://huggingface.co/standardmodelbio/SMB-v1_Qwen3-1.7b_multi-objective) | Qwen3 · 1.7B | Multi-objective |
| [SMB-v1_Qwen3-8b_multi-objective](https://huggingface.co/standardmodelbio/SMB-v1_Qwen3-8b_multi-objective) | Qwen3 · 8B | Multi-objective |
| [SMB-v1_Qwen3-8b_curriculum](https://huggingface.co/standardmodelbio/SMB-v1_Qwen3-8b_curriculum) | Qwen3 · 8B | Curriculum |
| [SMB-v1_Llama3-8b_multi-objective](https://huggingface.co/standardmodelbio/SMB-v1_Llama3-8b_multi-objective) | Llama3 · 8B | Multi-objective |
| [SMB-v1_Llama3-8b_curriculum](https://huggingface.co/standardmodelbio/SMB-v1_Llama3-8b_curriculum) | Llama3 · 8B | Curriculum |

## smb-encoders

Modality-specific encoders for 3D CT, somatic genomics, and proteomics. Drop in alongside smb-v1 for multimodal fusion, or use standalone.

| Encoder | Size | Modality |
|---|---|---|
| smb-vision-v1 | 0.6B params | CT |
| GenVarFormer | 993K params | Whole-genome sequencing |
| KRONOS | 7B params | Proteomics |

## Three ways to deploy the Standard Model encoders

1. **Linear probe · frozen embeddings** — freeze smb-v1, extract patient embeddings via last-token pooling, and train a lightweight head (logistic regression, Ridge, or CoxPH) to predict readmission, staging, treatment response, or survival. [Run the demo](https://standardmodel.bio/example.html)
2. **Multimodal fusion · projection layer** — combine longitudinal EHR embeddings from smb-v1 with modality-specific encoders for CT, somatic genomics, and proteomics. A learned projection layer maps each modality into a shared latent space, producing one patient representation that captures clinical, molecular, and radiological signal together.
3. **Multimodal alignment · LLM backbone** — align patient embeddings from smb-v1 and smb-encoders with external encoders or an LLM backbone for VQA, report generation, and other generative clinical tasks.

## Papers behind the weights

See [papers & research](https://standardmodel.bio/research.html) for the full list, including:

- The Patient is Not a Document: A World Model Training Paradigm for Oncology
- Predicting gene expression from long-range mutations in cancer (GenVarFormer)
- Patient-specific biomolecular instruction tuning of Graph-LLMs
- Validating the world model at Memorial Sloan Kettering (parts 1 and 2)

## Next steps

- [Install & quickstart](https://standardmodel.bio/install.html) — running locally in ~60 seconds
- [End-to-end example](https://standardmodel.bio/example.html)
- [Use it on your data](https://standardmodel.bio/your-data.html)
- Model access questions: info@standardmodel.bio

© 2026 Standard Model Biomedicine · San Francisco, CA