Prepare and test your skills
Prepare and test your skills
Gemini (opens in a new tab) is Google Cloud's family of large foundation models for language, code, and multimodal work. Gemini models can translate language, summarize text, generate code and creative writing, power chatbots and virtual assistants, and support search and recommendation systems. Teams use Gemini through Vertex AI (opens in a new tab) and related console tools to build content generation, reasoning, and coding applications.
Business uses for Gemini turn language understanding into content generation, sophisticated reasoning, and automated coding. Content generation covers translation, summarization, creative writing, chatbots and virtual assistants, and help for search and recommendation systems. Sophisticated reasoning handles complex or technical questions, while automated coding generates software code, including through Gemini Code Assist. Reasoning can be shallow on specialized topics that need deep domain expertise, and coding help can produce incorrect or inefficient code from edge cases, hallucinations, or weak prompts.
Teams improve results by tuning Gemini on domain-specific data and adding meaningful human supervision. Human review is especially important when output can affect a person's rights, finances, employment, or access to services. Users of Gemini Code Assist must understand its limits even though it is built with Google's AI principles.
Long-context work with Gemini means asking the model to consider a lot of information at once, within a fixed token limit. The maximum input and output token limit controls how much text Gemini can read in one request and how much it can write back. Teams must design large summarization, translation, and analysis jobs to fit inside that limit, often by splitting work into smaller parts. Grounding (opens in a new tab) improves accuracy by connecting output to real-world information and trusted business data, which reduces fluent but false or irrelevant answers.
Safety controls help enterprises deploy responsible AI (opens in a new tab) with Gemini. Vertex AI Studio provides built-in content filtering, while generative AI APIs provide safety attribute scoring that helps teams test filters and set confidence thresholds. Customers remain responsible for following Google Cloud's Acceptable Use Policy, Generative AI Prohibited Use Policy, and Service Specific Terms. Deployment limits that need extra testing include edge cases, hallucinations, poor prompts, training-data bias, lower quality for non-English languages and dialects, and limited depth in specialized domains.
Natively multimodal input means one Gemini model can accept text, images, video, audio, and documents such as PDFs in the same request. Because all parts reach the model together, the model can reason across them at the same time. For example, a prompt can include an image plus a video and ask when the image appears in the video and what the surrounding context means.
Each media part carries its own MIME type so Gemini knows whether it is receiving an image, video, audio clip, or document. A part can point to a file in a Cloud Storage bucket by its URI or carry a file uploaded from a local system. Each part can also carry a media resolution setting of MEDIA_RESOLUTION_HIGH or MEDIA_RESOLUTION_LOW, and a single request can mix levels for different files. PDFs default to a resolution of 560 when no value is set, so setting the correct MIME type and resolution matters for accurate interpretation.
PDF understanding improves when the request puts a single PDF before the text prompt so the model reads the document first. Long documents should be split into multiple PDFs so each part fits and processes cleanly. Machine-readable PDFs with rendered text work better than scanned images of text because the text is easier to edit, search, and manipulate. Spatial reasoning is still approximate, so Gemini may only estimate object locations and counts, and accuracy can drop on handwriting.
The Agent Studio workflow is the ordered path for building a multimodal prompt in the Google Cloud console. Agent Studio is the console page where the user creates the prompt, adds media, and runs text plus media together:
Media can come from several peer sources:
Advanced parameters shape generation. Top-K limits token choice to the K most probable tokens, Top-P selects from the most probable tokens until their probabilities sum to the chosen value, and Temperature controls randomness, where lower values give less open-ended answers and 1.0 is the recommended starting value. Other controls include the number of responses, streaming responses, the safety filter threshold, the output token limit, and up to five stop sequences. Grounding is not supported for multimodal prompts, so a multimodal request cannot be grounded the same way as a text-only prompt.
Exam tip: Grounding with specific data reduces hallucinations for text, but grounding is not supported for multimodal prompts in Agent Studio.
Search representations from Gemini Embedding 2 are numeric vectors for comparison and search rather than generated text. The model named gemini-embedding-2 accepts text, images, audio, and video as input only and produces embeddings as its only output. It supports output dimensions up to 3,072 through Matryoshka Representation Learning support, plus document OCR and audio track extraction from video. It accepts up to 8,192 input tokens and supports batch inference and standard pay-as-you-go, but not provisioned throughput or fixed quota.
Gemini model variants trade reasoning power against speed and cost. Gemini Ultra, Gemini Pro, and Gemini Flash all handle core tasks such as translation, summarization, code generation, chat, and support for search, while image-capable versions add multimodal work:
| Variant | Reasoning strength | Latency and cost focus | Choose when |
|---|---|---|---|
| Gemini Ultra | Highest for complex reasoning | Higher cost and latency trade-off | The task needs the deepest understanding and multi-step reasoning |
| Gemini Pro | Strong and balanced | Moderate latency and cost | The task needs solid quality for general business and developer work |
| Gemini Flash | Sufficient for common tasks | Lowest latency and most cost-effective | The task needs fast responses at high volume and low cost |
The Flash family includes Gemini Flash, Gemini Flash-Lite, and Gemini Flash Image versions such as gemini-2.5-flash and gemini-2.5-flash-lite. Gemini 2.5 Pro has a 10 queries per minute limit by default, while Gemini 2.0 Flash and Gemini 2.0 Flash-Lite use Standard pay-as-you-go with no default quota.
Continue with this study guide: Gemini
Gemma (opens in a new tab) is a family of small, lightweight open foundation models that teams deploy and serve inside Google Cloud. Unlike a fully managed model that only runs behind a vendor API, an open model can be self-deployed on infrastructure the customer controls. Gemma models are found in Model Garden and run on Gemini Enterprise Agent Platform inside the customer's project and virtual private cloud (VPC) network.
The self-deployment boundary defines who runs a Gemma model and where traffic stays. A self-deployable open model is a model the customer deploys rather than consumes as a hosted service. Self-deployed models are not serverless, so the customer must deploy them on Gemini Enterprise Agent Platform before use, and then the model, its endpoint, and the network path stay inside the customer's Google Cloud project and VPC network. To find them, open Model Garden and use the Features filter to select Open models and One-click deployment.
The same open model may be offered in two ways, and the model card tells which one is shown:
| Offering | Who runs the model | Where it runs | How you reach it |
|---|---|---|---|
| Managed API (MaaS) | Google-hosted service | API call | |
| Self-deployable open model | You | Your Google Cloud project and VPC network | Endpoint you deploy on Gemini Enterprise Agent Platform |
When a model is offered both ways, the managed API model card has API Service in its name, while the self-deployable card does not.
The Gemma model family shares a small, lightweight design but specializes by task from general text to code, safety, vision, and medical data. Choosing within the family means matching the specialty to the work:
Variant IDs in Model Garden identify the exact model, version, and size, such as google/gemma3@gemma-3-1b-it and google/gemma2@gemma-2-27b-it. The -it suffix marks instruction-tuned versions that follow prompts, while larger sizes such as 12B or 27B trade more compute for more capability than 1B and 2B models.
Exam tip: PaliGemma is the vision-language variant for image captioning, while numbered Gemma models are text models and -it marks instruction-tuned prompt-following versions.
Deployment of a Gemma model to an endpoint creates a serving endpoint that hosts the model and answers prediction requests. Model Garden verifies a deployment configuration for each model, so the same model can be deployed from the Google Cloud console or programmatically through the SDK, gcloud, REST, or Terraform. Deployment through Gemini Enterprise Agent Platform follows a fixed order:
model_garden.list_deployable_models or gcloud ai model-garden models list --model-filter=gemma, then record the model ID to deploy.machine_type="g2-standard-48", accelerator_type="NVIDIA_L4", and accelerator_count=4, and accept the end-user license agreement.After deployment, the endpoint is identified by a project ID, a region, and an endpoint ID. The endpoint ID can be found in the console under Online prediction > Endpoints or with gcloud ai endpoints list. Running inference sends a prompt to the endpoint path projects/PROJECT_ID/locations/ENDPOINT_REGION/endpoints/ENDPOINT_ID by calling predict. The sample configuration uses max_tokens of 1024, temperature of 0.9, top_p of 1.0, and top_k of 1 to control length and randomness.
Deployment paths differ for ABAP SDK (opens in a new tab) applications. The Agent Platform path uses class /GOOG/CL_GEMMA_ON_VERTEXAI with model key, project number, region, and endpoint ID, while the Cloud Run or Gemini API path uses class /GOOG/CL_GEMMA_ON_CLOUDRUN with only the model key. Before sending requests, teams enable required APIs, deploy the model if needed, install the SDK, set up authentication, and configure service mapping.
Serving containers and tuning for Gemma determine where the model runs and how its behavior is adjusted. Different serving containers support different workloads on Cloud TPUs and Cloud GPUs, while Workbench fine-tuning notebooks adjust behavior for a specific use case. Serving can also use Spot VMs or a Compute Engine reservation to control cost and capacity. Popular serving frameworks include vLLM, Hex-LLM, SGLang, TGI, and TensorRT-LLM, and a custom vLLM container is used when the model needs unique pre-processing, post-processing, dependencies, or configuration.
Finding Gemma starts in Vertex AI Model Garden, the catalog of Google, partner, and open-source models for discovery, testing, customization, and deployment. Model Garden organizes models into foundation models, fine-tunable models, and task-specific solutions, with consistent patterns for tuning, evaluation, and serving. Gemma can also be used with Google Kubernetes Engine (GKE) (opens in a new tab), Dataflow, and Colab.
Responsible controls combine scanning, access limits, transparency, and evaluation. Model Garden tests serving and tuning containers, scans container artifacts for vulnerabilities, and scans third-party checkpoints for authenticity, malware, pickle files, Keras Lambda layers, and secrets. A Model Garden organization policy set at the organization, folder, or project level can allow only vetted models and deny all others. Models deemed unsafe are blocked from deployment, while models deemed suspicious or capable of executing remote code are only flagged and can still be deployed, so thorough review is needed before use.
Continue with this study guide: Gemma
Imagen (opens in a new tab) is Google Cloud's foundation model for high-quality image generation and editing. Imagen creates new high-fidelity images from text prompts and transforms existing images with guided editing. The core flow moves from a text description to a new image, and from a source image plus a text description to an edited image.
Text-to-image generation is Imagen's primary function for creating original images from a written prompt. A user provides a text description of the desired scene, object, or style, and Imagen generates new image content that matches it. This supports creative production, marketing visuals, and concept exploration without starting from an existing photo, and clearer prompts produce better results.
Inpainting edits content inside the boundaries of an existing image while preserving the rest of the picture. The source image stays as the base, and Imagen regenerates only the targeted area according to the prompt:
Outpainting expands content beyond the original edges by generating a natural continuation of the scene. Image-to-image transformation is the broader pattern behind all edits, where an input image plus a text prompt produces a new output image.
Exam tip: Inpainting edits content inside the image, such as inserting or removing an object, while outpainting expands content outside the original image boundaries.
Business capabilities for Imagen are the Generally Available image functions that map to different content tasks. Text-to-image generation creates original assets, image editing modifies an existing image with a prompt, image visual captioning generates a text description of an image, and Visual Question Answering (VQA) answers questions about image content. Subject model tuning and style model tuning customize output to a specific subject or visual style. Human face generation is enabled by default except for images with children and celebrities.
Industry applications center on automated marketing asset creation and product visualization. Marketing teams use generation to produce campaign visuals, banners, and variations quickly from prompts instead of manual photoshoots. Product teams use editing plus subject and style tuning to show the same product in new colors, settings, or brand styles while keeping its identity consistent. Captioning and VQA support catalogs, accessibility, and customer experience by turning images into searchable text that flows from Imagen to a search or storefront system.
Enterprise deployment runs on fully managed Vertex AI infrastructure rather than customer-managed servers. The application sends a prompt or image request to the Imagen API on Vertex AI, Vertex AI runs inference, and the image or text returns to the application. Vertex AI Inference supports online inference for real-time individual requests and batch inference for groups of requests returned together. Regional APIs let customers control where data is stored at rest, and Vertex AI Model Garden provides centralized discovery of Google and third-party models.
Choosing Imagen depends on the required output type, since Google Cloud routes generative tasks by output:
A decision tree branching on required output type to pick Imagen for images, Veo for video, Codey for code, text-embedding for vectors, and Gemini for text from multimodal input.
| Task | Model to use | Choose when |
|---|---|---|
| Generating images | Imagen | The output must be a new image or an edited image |
| Generating code | Codey | The output must be code completion, generation, or chat |
| Generating embeddings | text-embedding | The output must be vectors for search, classification, or clustering |
| Generating text with images or video in the prompt | Gemini | The input is multimodal and the output is text |
Enterprise customization in Imagen lets organizations shape output to their own brand or subject matter without starting over. Editing includes inserting objects with inpaint, removing objects with inpaint, expanding content with outpaint, and replacing backgrounds, plus Virtual Try-On images for retail and ecommerce. Customization paths include subject customization, style customization, controlled customization, instruct customization that follows natural-language instructions, and upscaling that increases resolution.
Responsible AI safety controls are configurable settings that let administrators decide how Imagen filters content. Responsible AI and usage guidelines define acceptable use, while a prompt rewriter refines user prompts and negative prompts omit unwanted content. Filtering is therefore an organizational decision, not a single fixed behavior. Production parameters control delivery, including aspect ratio and output resolution for shape and size, text prompt language, and deterministic generation for repeatable results when the same prompt must produce consistent output.
Continue with this study guide: Imagen
Veo is Google's high-definition video generation model for creative storytelling and visual communication. Veo turns text prompts and images into short, cinematic MP4 clips. Teams access Veo in the Gemini Enterprise Agent Platform Media Studio console or through the Veo API, so non-specialists and developers can create video from the same model.
Creative building blocks in Veo are the generation and editing capabilities that carry a project from idea to revision. A text prompt, still image, or existing clip flows into Veo as input, and Veo returns an MP4 video clip as output:
Veo is built for short-form output assembled from shots rather than full-length films in one step. A single generation produces clips that are 4, 6, or 8 seconds long at 24 frames per second, with up to 4 variations per prompt. Supported formats are 16:9 landscape and 9:16 portrait at 720p or 1080p resolution. Veo 3.1 Preview models add a 4K option, while Veo 3.0 tops out at 1080p.
High-impact marketing with Veo is rapid creation of short, polished spots, product teasers, and social clips without a full shoot. A marketer describes the scene, product, and cinematic style in text, and Veo can generate up to four options to compare. An existing product photo can serve as the first frame, so data flows from brand-owned imagery into Veo and out as motion that preserves the product look. Reference images keep a product, mascot, or style consistent, while edit and extend functions revise scenes or stretch strong shots. Sound generation plus landscape and portrait ratios let one concept adapt to television-style spots and vertical mobile feeds.
Internal communications with Veo uses short custom video to explain ideas, announce changes, and train employees more clearly than text alone. A leader turns a script into a visual scene such as a walkthrough or welcoming message, often by animating a company slide, diagram, or workplace photo as the first frame. Teams can generate multiple tone or style variations from one prompt and pick the clearest version. Short 4- to 8-second clips work well as town-hall intros, training segments, or highlights in company news.
Cinematic prototypes with Veo are fast pre-visualizations for media and entertainment teams. A director describes camera movement, lighting, and action, and Veo returns a clip that shows how the idea could look on screen. First-and-last-frame control defines opening and closing compositions while Veo generates the connecting motion, and reference images guide characters, props, and art direction. Edit and extend functions then revise scenes or lengthen promising shots, with Veo 3.1 adding higher-resolution drafts.
Exam tip: Veo generates short clips of 4, 6, or 8 seconds, up to 4 per prompt, so choose Veo for ads, explainers, and prototypes assembled from short shots.
Video generation modes in Veo are defined by what the user supplies, but not every version supports every mode. Veo 3.0 supports text to video and image to video but does not support extending videos or using reference images. First-and-last-frame generation is supported by veo-3.1-generate-001, veo-3.1-fast-generate-001, veo-3.1-lite-generate-001 in preview, and gemini-omni-1.1-flash-preview.
Technical specifications set fixed limits for duration, count, size, shape, resolution, frame rate, and file type. Veo 3.0 allows 4, 6, or 8 seconds with a default of 8 seconds, up to 4 videos per prompt, image inputs up to 20 MB as image/jpeg or image/png, 9:16 portrait and 16:9 landscape with 16:9 as default, 720p and 1080p output, 24 FPS, and video/mp4 output. Veo 2 accepts 5 to 8 seconds with a default of 8 seconds, while Veo 3.1 Preview models add 4K with a default of 720p.
A Veo generation request works as a long-running operation that writes video to storage or returns bytes directly:
Prompt controls refine framing, content, safety, and reproducibility. aspectRatio sets 16:9 or 9:16, negativePrompt describes content to avoid, personGeneration controls people and faces with allow_adult permitting only adults, resolution applies to Veo 3 models only, sampleCount sets 1 to 4 videos, and seed fixes the starting point for repeatable results. Availability is in the United States region us-central1 with English prompts, with 10 tokens per minute for veo-3.0-generate-001 and 50 tokens per minute for veo-3.1-lite-generate-001. Provisioned Throughput and fixed quota are supported, while batch inference and pay-as-you-go are not, and generated content carries Content Credentials.
Supporting services for video work connect Veo-style creative output to secure delivery, observability, and key management. Video delivery can use Google Cloud Video Stitcher API to register livestream events for server-side ad insertion, with traffic flowing from the Video Stitcher API to the Interactive Media Ads Dynamic Ad Insertion SDK on the player. Setup links the cloud project, live event, and ad serving through location, project number, OAuth token, network code, live config ID, and custom asset key.
Monitoring health uses Cloud Monitoring for integrations, dashboards, and alerting policies:
The policy depends on the notification channel to deliver the alert, so a policy with disabled channels still fires but fires silently. Encryption coverage uses Cloud Key Management Service (Cloud KMS) to show how customer-managed keys protect data at rest by protection type, by service, and by alignment to key practices. Viewing metrics requires the Encryption Dashboard Viewer role, while viewing customer-managed encryption key coverage details also requires the Cloud Asset Viewer role.
Continue with this study guide: Veo
Gemini Ultra is for tasks needing the deepest understanding and multi-step reasoning despite higher cost and latency, Gemini Pro is for general business and developer work needing solid balanced quality at moderate latency and cost, and Gemini Flash is for common tasks needing fast responses at high volume and low cost with the lowest latency. All three handle core tasks such as translation, summarization, code generation, chat, and support for search, while image-capable versions add multimodal work.
Self-deployable Gemma open models are deployed and run by the customer inside the customer's Google Cloud project and virtual private cloud (VPC) network on Gemini Enterprise Agent Platform and are reached through an endpoint the customer deploys. Managed API models are run by Google on a Google-hosted service and are reached by API call.
Inpainting in Imagen edits content inside the boundaries of an existing image while preserving the rest of the picture, such as inserting objects, removing objects, replacing the background, or editing details through guided prompts. Outpainting expands content beyond the original edges by generating a natural continuation of the scene.
Veo turns text prompts and images into short cinematic MP4 clips, including text to video, first frame image to video, first and last frames to video, generation from references, extension, and editing. A single generation produces clips that are 4, 6, or 8 seconds long at 24 frames per second with up to 4 variations per prompt through a long-running operation that writes video to Cloud Storage or returns video bytes directly.
Last updated