Model Garden on Google Cloud: Model as a Service vs. Self-Deployment

GCP Study Hub
November 15, 2026

Model Garden on Google Cloud offers two ways to run models: Model as a Service (MaaS) and self-deployment. Choosing between them on the Professional Agentic Architect exam comes down to whether a workload can use a standard model on a shared managed API or requires custom weights, specific GPU hardware, predictable high-volume serving, or isolation inside your own project and VPC network.

Model as a Service (MaaS)

Model as a Service is a serverless, multi-tenant managed API for standard models. It requires the least operational work when a standard model meets the application's needs and using a shared service is acceptable.

That convenience comes with clear limits. With MaaS, prompts are sent to a shared, multi-tenant managed service, and you cannot bring custom weights. When regulations prohibit sending prompts to a multi-tenant managed service, or when an application relies on a fine-tuned model with custom weights, MaaS cannot meet those requirements.

Self-deployment to dedicated endpoints

Self-deployment runs a model on a dedicated endpoint inside your own Google Cloud project. Google recommends self-deployment in several specific situations:

  • You need to serve custom weights or a fine-tuned version of a model rather than a standard managed model.
  • The workload has predictable, high-volume traffic.
  • You need to choose exact machine types and GPUs, such as the NVIDIA L4.
  • The model must stay inside your own project and VPC network, including when regulations prohibit sending prompts to a multi-tenant managed service.

Using dedicated endpoints does not mean a team has to build its own serving stack. Model Garden provides one-click deployment and prebuilt serving containers such as vLLM, so teams can host models in their own project without maintaining custom serving infrastructure.

Distinguishing Model Garden endpoints from related services

Model hosting decisions on the Professional Agentic Architect exam also involve separating Model Garden endpoints from other components in an agent architecture.

Agent Runtime hosts agents, not GPU model endpoints. When a workload requires a model serving endpoint running on GPUs, it belongs on a self-deployed Model Garden endpoint rather than on Agent Runtime.

Switching a fine-tuned model to a managed model such as Gemini Flash throws away the fine-tuning. If a team has fine-tuned a model for its workload, moving to a standard managed model discards that work instead of serving the custom weights.

Matching the deployment model to workload constraints

In practice, the decision splits cleanly along operational simplicity versus control. MaaS fits standard models where a multi-tenant managed API is acceptable and minimizing operational work is the priority. Self-deployment through Model Garden fits workloads that require custom weights, dedicated project and VPC isolation, specific GPU choices such as the NVIDIA L4, or predictable high-volume serving.

Our Professional Agentic Architect practice exams test Model Garden deployment options alongside Agent Runtime and managed model selection, with detailed explanations that drill these distinctions.

Get tips and updates from GCP Study Hub

arrow