Loading…
Cloud catalog
Google Cloud Model Garden
Managed cloud endpoints for deploying and fine-tuning biological foundation models, billed per token or per node-hour depending on serving mode.
- Models served
- 8
- Offers
- Inference
- Coverage
- Imaging 4Pathology 2Language model 2DNA & Gene 1+2 more
- Deployment
- Managed GCP · your project
- Pricing
- Per token or per node-hour
Models you can run on Google Cloud Model Garden
Most cited first42533.2K305
Medically tuned SigLIP encoder from Google that maps medical images and text into one embedding space for zero-shot classification and retrieval.
ImagingPathology
425131.8K1.6K
Open medical multimodal models from Google, built on Gemma 3 with a medically tuned SigLIP vision encoder for clinical text and image understanding.
Language modelImaging
206—860
DNA foundation model that predicts thousands of functional genomic tracks, from expression and splicing to chromatin, at single base-pair resolution.
DNA & Gene
685838
Open therapeutics foundation models from Google, built on Gemma-2, for drug-discovery property prediction and conversational reasoning.
Language modelSmall molecule
4121200
Chest X-ray embedding model built on ELIXR, producing image and image-text embeddings for data-efficient and zero-shot radiograph classification.
Imaging
How access works
Access requires a Google Cloud project with billing enabled and standard platform credentials; there is no separate sign-up. New accounts receive free credits, usually enough to evaluate a model before committing. Three consumption shapes behave differently: first-party and partner models are consumed as managed APIs with nothing to deploy, while open models are self-deployed to an endpoint in the customer's own project. A HIPAA business associate agreement covers the hosting platform.
- Self-serve
- Trial credits
- Public price list
- API access
- SSO / SAML
- HIPAA BAA
- Region selectable
What it costs
Managed models bill per token, with separate rates for input, output and cached input. Open models deployed to an endpoint bill for the compute behind it instead, per node-hour per node. That second mode carries the cost trap: billing applies to every model deployed to an endpoint whether or not it serves a prediction, because a node-hour covers idle time in an active state as well as serving time. Undeploying stops the meter; ceasing to call it does not. New accounts start with free credits.
Last verified
