
Adopting a first-party, open-weight, or partner model on Google Cloud begins with finding the right model and then deciding how much of the serving stack to manage. Inside Gemini Enterprise Agent Platform (formerly Vertex AI), Model Garden acts as that catalog layer, providing a single AI and ML model library to discover, test, customize, and deploy models from Google and its partners. Across Professional Machine Learning Engineer scenarios, the main architectural decision is matching requirements to the right serving path: calling a model through Model as a Service (MaaS) with no infrastructure to manage, deploying an open model to a managed Agent Platform endpoint, hosting it on Google Kubernetes Engine (GKE) for full infrastructure control, or packaging a custom Docker container when specialized inference logic is needed.
Model Garden brings first-party, open-weight, and partner models into one catalog. Teams can browse the library visually in the Google Cloud console or integrate catalog discovery and deployment into code through the Model Garden SDK, the API, and the CLI.
Once a team identifies a model in Model Garden, the next decision is whether that model will be called directly as a serverless API or hosted on infrastructure that the team deploys.
Model as a Service, or MaaS, is a serving model rather than a resource that you deploy or a separate product you navigate to in the console. It serves selected models through a serverless API so that Google Cloud hosts the model and applications simply send requests to it. Because MaaS handles hosting behind the scenes, there is no endpoint to create, no machine type or accelerator to select, and no autoscaling to configure.
Using a MaaS model involves two specific steps. First, you must enable that model's API in Model Garden before calling it. Each MaaS model has its own API to turn on, so access is granted per model rather than enabled across the entire catalog at once. Second, once enabled, applications call the model through the Google GenAI SDK for Python by referencing its model ID. The application code stays the same across models, with the SDK handling the request and the model ID determining which MaaS model generates the response.
When a team hosts an open model from Model Garden rather than consuming a serverless MaaS API, there are two hosting paths. The choice depends on how much of the serving stack the team wants to own.
Many models in Model Garden support one-click deployments to a fully managed Agent Platform endpoint. You select the model and configure the deployment settings, and Model Garden provisions the endpoint for you. This path delivers faster deployments along with the built-in MLOps capabilities of Agent Platform serving.
The alternative is deploying the model to GKE, where it runs on a cluster your team manages. With GKE, you configure the nodes, accelerators, and scaling yourself. That adds setup and operational work, but it gives you full control over the underlying infrastructure. On Professional Machine Learning Engineer scenarios, we would lean toward GKE when strict infrastructure or compliance requirements cannot be met by a managed endpoint, and toward Agent Platform endpoints when rapid deployment and managed MLOps serving features are the priority.
One-click deployments rely on pre-built serving containers provided by Agent Platform. Those pre-built containers expect input data to arrive already formatted for the model and return the model output as is. When a use case requires specialized preprocessing or post-processing, such as custom data transformations, custom feature engineering, or framework-specific logic, the pre-built containers cannot support it.
To add custom inference logic without giving up managed serving, you package the model artifacts together with your custom inference code in a custom Docker container. The workflow follows five steps:
Even with those extra packaging steps, you retain the managed serving infrastructure of Agent Platform. The endpoint provides autoscaling based on traffic, monitoring, health checks, versioning, load balancing, and GPU support when required, all while running your custom preprocessing and post-processing code.
Across Professional Machine Learning Engineer architecture questions, each serving option from Model Garden maps to a distinct operational boundary:
Our Professional Machine Learning Engineer course covers Agent Platform Model Garden and Model as a Service alongside custom inference containers and model serving architectures, with practice questions that drill these distinctions.