Model Garden is the model library of the Gemini Enterprise Agent Platform. It brings together models from Google, from partners, and from the open source ecosystem in one place and provides a consistent deployment pattern for all of them.
What is Model Garden?
Google describes Model Garden as an AI/ML model library that helps you discover, test, customize, and deploy models and assets from Google and Google partners. In the product catalog, Model Garden is listed as a single place to discover over 200 models from Google and Google partners.
The documented advantages are that all available models are grouped in a single location, that a consistent deployment pattern applies across different model types, that there is built-in integration with other parts of the Gemini Enterprise Agent Platform such as model tuning, evaluation, and serving, and that the platform handles deployment and serving of generative AI models.
Models are grouped into three categories: foundation models as pretrained multitask large models that can be tuned or customized using Agent Studio, the Agent Platform API, and the SDK; fine-tunable models that can be fine-tuned using a custom notebook or pipeline; and task-specific solutions, most of which are prebuilt and ready to use and many of which can be customized with your own data. The filter pane narrows results by tasks, model collections, providers, and features; details are on each model card.
On security, Google describes thorough testing and benchmarking of the serving and tuning containers it provides, plus active vulnerability scanning of container artifacts. Third-party models from featured partners undergo model checkpoint scans to ensure authenticity; models from HuggingFace Hub are scanned by HuggingFace and its third-party scanner for malware, pickle files, Keras Lambda layers, and secrets. Models deemed unsafe are flagged and blocked from deployment, while models deemed suspicious are indicated but can still be deployed.
For open source models, the documentation names three cost items: model tuning at the same rate as custom training, model deployment charged for the compute resources of the endpoint per predictions pricing, and Colab Enterprise per its own pricing.
Core Features
- Central model catalog: Models from Google, partners, and the open source ecosystem in one place.
- Three model categories: Foundation models, fine-tunable models, and task-specific solutions.
- Filters and model cards: Narrowing by tasks, model collections, providers, and features with per-model details.
- Consistent deployment: One deployment pattern across different model types.
- Platform integration: Connection to tuning, evaluation, and serving on the Gemini Enterprise Agent Platform.
- Model security scanning: Container scans, checkpoint scans for partner models, and screening of HuggingFace models.
- Governance: Organization policy to restrict access to specific models.
Typical Use Cases
Selecting a model for a new application
A team compares several foundation models using model cards and filters before deciding on one variant.
Customizing an open model
An open model is fine-tuned through a notebook and then deployed to an endpoint.
Operating without your own serving infrastructure
Instead of building serving infrastructure, the platform handles deployment and serving of the model.
Governance over usable models
An organization restricts which models may be used in projects through the organization policy.
Reviewing third-party models
Before using a model from HuggingFace Hub, the scan results and indicators described by Google are evaluated.
Benefits
- One catalog instead of many sources: Google, partner, and open models are discoverable together.
- Consistent path to deployment: The same pattern for different model types.
- Platform integration: Tuning, evaluation, and serving work together directly.
- Transparency on model security: Scanning and screening mechanisms are documented.
- Controllable: Access to models can be restricted through an organization policy.
Integration with innFactory
As a certified Google Cloud Partner, innFactory supports you with Model Garden: selecting suitable models for your use case, assessing cost and regional availability, governance concepts for model access, and building tuning and serving processes. We describe what a company-wide AI stack can look like using our CompanyGPT cloud stack as an example.
Contact us for a consultation on Model Garden and Google Cloud.
Typical Use Cases
Technical Specifications
Frequently Asked Questions
What is Model Garden?
According to the documentation, Model Garden is an AI/ML model library that helps you discover, test, customize, and deploy models and assets from Google and Google partners. Google's product catalog describes Model Garden as a single place to discover over 200 models from Google and Google partners.
Which model categories exist?
The documentation distinguishes three categories: foundation models, which are pretrained multitask large models that can be tuned or customized using Agent Studio, the Agent Platform API, and the Agent Platform SDK; fine-tunable models, which you can fine-tune using a custom notebook or pipeline; and task-specific solutions, most of which are prebuilt and ready to use and many of which can be customized with your own data.
How do I find a suitable model?
The filter pane lets you filter by tasks, model collections - models managed by Google, partners, or you - providers, and features. Details on each model are available on its model card.
How are models from Model Garden billed?
For open source models in Model Garden, the documentation names three cost items on the Gemini Enterprise Agent Platform: model tuning is charged for the compute resources used at the same rate as custom training, model deployment is charged for the compute resources used to deploy the model to an endpoint per predictions pricing, and Colab Enterprise follows its own pricing.
How does Google handle model security?
Google performs thorough testing and benchmarking on the serving and tuning containers it provides and applies active vulnerability scanning to container artifacts. Third-party models from featured partners undergo model checkpoint scans to ensure authenticity. Models from HuggingFace Hub are scanned directly by HuggingFace and its third-party scanner for malware, pickle files, Keras Lambda layers, and secrets; models deemed unsafe are flagged and blocked from deployment in Model Garden. Models deemed suspicious are indicated but can still be deployed - Google recommends a thorough review in that case.
Can I restrict access to specific models?
Yes. The documentation describes a Model Garden organization policy that lets you control access to specific models.
Note: All product information on this page has been compiled with care, but is provided without guarantee and may be outdated or incomplete. Cloud services evolve rapidly — features, pricing, SLAs, and availability change frequently. Authoritative and up-to-date information can only be found on the official product page of Google Cloud (official documentation). This page does not represent an offer by Google Cloud.
