Prepare and test your skills

Prepare and test your skills

Pre-trained APIs are ready-to-use AI services that add machine learning (opens in a new tab) capabilities to an application without building or training a model. These models are already trained on vast datasets, so developers do not need specialized machine learning expertise to use them. Because training work is already done, pre-trained APIs reduce cost and development time and let a team move quickly from an idea to a working feature.
What pre-trained APIs provide is direct access to powerful AI capabilities through a simple request and response. A developer sends input such as an image, a block of text, an audio file, or a document to the API and receives structured results back. There is no need to collect training data, choose an algorithm, or manage training compute. This makes pre-trained APIs the fastest path when the task is common and already handled by an existing API.
Google Cloud pre-trained APIs are a set of specialized services where each API handles a specific type of input and task. A single application can combine several APIs into one workflow, such as sending an uploaded document to one API for extraction and then sending the extracted text to another API for analysis. The choice of API depends on whether the input is an image, text, audio, video, document, or conversation.
A taxonomy tree of Google Cloud pre-trained APIs, grouping each API by the type of input it handles: images, text, audio, video, documents, conversations, and general-purpose tasks, with typical use cases under each.
| API | What it does | Typical use cases |
|---|---|---|
| Cloud Vision API | Derives insights from images | Image analysis, content moderation, and automated data entry |
| Cloud Natural Language API | Analyzes text structure and meaning | Customer feedback analysis, content categorization, and social media trends |
| Speech-to-Text API | Converts audio to text | Transcription across many languages and dialects |
| Video Intelligence API | Analyzes video to identify objects, scenes, and actions | Video content analysis, moderation, and search |
| Document AI API | Extracts, classifies, and understands data in documents | Automating document processing workflows |
| Dialogflow API | Builds conversational interfaces such as chatbots and voice assistants | Customer service bots and virtual assistants |
| Gemini API | Provides access to Google's most capable general-purpose AI model | General-purpose AI tasks |
The Cloud Natural Language API is a natural language understanding service that answers different questions about a piece of text. The Cloud Natural Language API is part of the larger Cloud Machine Learning API family and offers five distinct analysis types. Language support varies by type, so supported languages should be checked before depending on a specific language.
Pre-trained APIs compared with custom model paths differ in speed, expertise required, and fit for unique data. Pre-trained APIs are the right starting point when the task is common and must ship quickly without machine learning expertise. When an organization needs a model tuned to its own categories, the alternative is a custom model such as AutoML Natural Language for content classification or a model from Model Garden. Model Garden groups pre-trained models into Google models like the Gemini family and Imagen, open models like Gemma and Llama, and third-party models from partners such as Anthropic.
A decision tree branching on task commonality, proprietary data, and control needs to choose pre-trained APIs, AutoML with BigQuery ML, fine-tuning, or custom training.
Exam tip: Choose a pre-trained API when a common task must ship quickly without machine learning expertise; choose AutoML or another custom path when the model must learn from your own labels or data.
Fine-tuning is a middle option between using an API as-is and training from scratch. Fine-tuning uses learning transfer, where knowledge from a pre-trained model is applied to a new task instead of starting over. Learning transfer requires less data and compute time and therefore lowers cost.
Custom model development and selection criteria compare training your own model against reusing managed models and APIs. Custom development means training and running your own model with full control over data, code, and infrastructure. Managed services mean reusing models and APIs that Google Cloud already provides. The best choice depends on data volume and compute, the need for total model control, and specific performance goals.
Starting with pre-trained models and managed APIs is the recommended path because training complex models from scratch requires significant time and computational resources. Pre-trained models are already trained on vast datasets, so they eliminate training from scratch and reduce cost and development time. Model Garden offers pre-trained models for image classification, object detection, and natural language processing, grouped into Google models such as Gemini and Imagen for image generation, open models such as Gemma and Llama, and third-party models from partners such as Anthropic. Managed APIs let developers integrate AI without building models, including Cloud Vision API for images, Cloud Natural Language API for text meaning, Speech-to-Text API for audio transcription, Video Intelligence API for video, Document AI API for documents, Dialogflow API for chatbots, and Gemini API for general-purpose tasks. Fine-tuning sits between reuse and custom building because learning transfer applies knowledge from a pre-trained model to a new task with less data and compute.
When custom model development makes sense is when pre-trained models cannot meet requirements for control, data fit, or performance. Custom development packages training code and dependencies into custom containers stored in Artifact Registry, which creates consistent and reproducible training environments across pipelines and projects. Trained models are then stored in Model Registry, which manages and deploys models so they can be reused in later pipelines or for online prediction. Google Kubernetes Engine (GKE) (opens in a new tab) is the platform choice when the team needs deep control, portability, and a customized high-performance AI platform:
GKE supports GPUs and TPUs for inference, the GKE Inference Gateway for model-aware routing and load balancing, faster startup nodes that reduce GPU node startup times by up to 80%, Kueue for job queueing and quota management, and TPU multislice, which lets multiple TPU slices communicate over the data center network for large-scale training.
Comparing managed and self-managed platforms shows a spectrum from fully managed to fully configurable infrastructure. The right platform depends on how much control, flexibility, and management the team wants to keep. Vertex AI offloads infrastructure management for rapid development, GKE keeps deep control and portability, and Cloud Run removes server management for containerized inference.
| Google Cloud service | Best for |
|---|---|
| Vertex AI | Fully managed, end-to-end platform that accelerates development and suits teams focused on MLOps and rapid time-to-value |
| GKE | Deep control, portability, custom pipelines, kernel-level customization, very large model training, and minimal-latency serving |
| Cloud Run (opens in a new tab) | Serverless containerized inference that scales to zero and suits event-driven applications and smaller models |
Cloud Run hosts AI applications by deploying application code or a container as a scalable API endpoint that handles incoming requests. The application sends data to a model for processing and returns results, where the model can be a Google model such as Gemini or a Vertex AI model, an open-source model such as Llama or Gemma, or a custom trained model. Cloud Run GPUs support computationally demanding inference workloads that require a large number of operations.
Selection criteria and trade-offs center on data volume, the need for total model control, and specific performance objectives. Testing hypotheses at small scale with smaller models and datasets reduces cost, allows faster iteration, simplifies debugging, and avoids over-provisioning. Teams should start with a representative data sample, experiment in notebooks such as Agent Platform Workbench or Colab Enterprise, and use BigQuery ML for structured data, where models are created and deployed directly within BigQuery and billed per query. Reuse also lowers cost because Feature Store centralizes ML features for reuse across projects and models.
Exam tip: Fine-tuning a pre-trained model reuses existing knowledge and needs less data and compute, while training from scratch requires significant resources and time, so start with pre-trained models and choose custom training only when control or performance demands it.
Analyze the role of AutoML in custom solutions means understanding how teams build high-quality models from proprietary data without a large data science team. AutoML is the middle path between using a ready-made pre-trained API and building a fully custom model with Python or Java. BigQuery ML (opens in a new tab) is an example of this approach because SQL practitioners can create and run machine learning models directly where their data already lives.
Custom models built from proprietary data are models that learn from training data stored in an organization's own datasets. A model in BigQuery ML represents what a machine learning system has learned from training data, and BigQuery ML models are stored in BigQuery datasets similar to tables and views. BigQuery ML also connects custom models to ready-made intelligence by letting models access Gemini Enterprise Agent Platform models and Cloud AI APIs for tasks like text generation or translation. In practice, unique business data such as sales history, customer behavior, or time series metrics becomes a reusable model object inside the same warehouse that holds the data.
Lowering the expertise needed to build models is the main advantage of AutoML for business teams. AutoML lets data analysts use existing SQL tools and skills to build and evaluate models, including default settings in CREATE MODEL statements and inference functions. This matters because data analysts understand the data but often have limited machine learning and programming expertise, while traditional development restricts work to a small set of specialists. Basic knowledge of feature engineering and model training still helps optimize data and models for better results. Teams can work with BigQuery ML through familiar entry points:
bq command-line toolFaster time to market by bringing ML to data happens because large datasets do not need to move to a separate machine learning environment. BigQuery ML increases development speed because moving and formatting large amounts of data for Python-based frameworks is not required. That approach reduces complexity because fewer tools are required and helps teams reach production faster. Predictive analytics built this way can guide business decisions across the organization, including forecasting and anomaly detection on time series data and contribution analysis for changes in metrics like revenue.
Exam tip: BigQuery ML creates custom models with CREATE MODEL using default settings, while Cloud AI APIs perform ready-made tasks like translation, so choose the custom-model path when predictions must learn from your own proprietary dataset.
CREATE MODEL, avoiding the need to move large datasets to a separate environment.Use Google Cloud pre-trained APIs when the task is common and must ship quickly without machine learning expertise, because the models are already trained on vast datasets and work through a simple request and response. Choose AutoML or another custom path when the model must learn from your own labels or data, such as AutoML Natural Language for content classification.
The Cloud Natural Language API provides five distinct analysis types for understanding text structure and meaning. Sentiment analysis measures the overall emotional tone, entity analysis identifies the people, places, and things mentioned, entity sentiment analysis shows the sentiment attached to each identified entity, content classification sorts text into categories, and syntax analysis breaks down the grammatical structure.
BigQuery ML lets SQL practitioners create and run custom models directly where their data already lives inside BigQuery, so large datasets do not need to move to a separate machine learning environment. A model represents what the system has learned from proprietary training data such as sales history or customer behavior and is stored in BigQuery datasets similar to tables and views, which reduces complexity and helps teams reach production faster.
An enterprise is transitioning an open large language model (LLM) from an initial prototype to full production on Google Cloud. The engineering team has established the following technical and operational requirements for the deployment:
Which Google Cloud model serving approach should the enterprise choose?
Self-deployment using a custom vLLM container
Model as a Service (MaaS) serverless managed APIs
Gemini Enterprise Agent Platform prebuilt container images
One-click deployment of a standard model in Model Garden
Last updated