Task 1.4: Identify the use cases and strengths of Google's foundation models.

• AGemini, called through Agent Platform

  Incorrect

  Gemini is offered as a managed model through an API; the hospital would not host its weights in its own environment.

• BVeo

  Incorrect

  Veo is not a language model for the hospital's text tasks.

• CGemini Nano Banana image models

  Incorrect

  Gemini's image models generate and edit images and are offered as managed models, not as open weights to run in-house.

• DGemma

  Correct

  Gemma models are Google's open models: you can download them, deploy them on your own infrastructure or endpoints, and tune them for your use case.

Open weights, self-hosting and full customization point to Gemma; managed access with no infrastructure to run points to Gemini.

• AGemma

  Incorrect

  Gemma models are open models for tasks such as text generation, coding and reasoning; they do not generate video.

• BVeo

  Correct

  Veo generates video from text prompts and from images, including a first frame image, which matches clips built from a description and a product photo.

• CGemini 3.1 Flash-Lite Image

  Incorrect

  Gemini's image models create still images; the brand needs moving video clips.

• DGemini Embedding

  Incorrect

  Embedding models turn content into vectors for search and similarity; they do not generate any media.

Video generation from text or images is Veo's job; match the output modality first when choosing among Google's models.

• AVeo 3.1

  Incorrect

  Veo is not built for refining still images over several turns.

• BThe Cloud Vision API for image analysis

  Incorrect

  The Vision API analyzes existing images; it does not create or edit them.

• CNano Banana 2, a Gemini image model

  Correct

  Gemini's image generation models support conversational, multi-turn editing and character consistency across images, which is the workflow described.

• DGemma 3, an open multimodal model

  Incorrect

  Gemma 3 accepts text and image input but produces text; it does not generate images.

Image generation and conversational editing belong to Gemini's image models; analysis-only APIs and text-output models do not produce images.

• AVeo

  Incorrect

  Veo does not answer questions about existing footage.

• BGemma 2

  Incorrect

  Gemma 2 is an open text model for generation, summarization and extraction; it cannot take in the recorded video.

• CGemini

  Correct

  Gemini models are multimodal: they understand text, images, audio and video in the same request and reason across them, with large context windows for long inputs.

• DGemini Embedding

  Incorrect

  Embedding models produce vectors for search; they do not answer questions about interviews.

Understanding several modalities at once, at length, is Gemini's strength; Veo generates media rather than understanding it.

• AGemma 3n

  Correct

  Gemma 3n is an open model designed for efficient execution on low-resource devices, with multimodal input including text, image, video and audio.

• BGemini 2.5 Pro

  Incorrect

  Gemini 2.5 Pro is a high-capability managed model reached through an API; it does not run offline on a phone.

• CVeo 3.1

  Incorrect

  Veo generates video; it is not a small on-device model for an app.

• DGemma 2

  Incorrect

  Gemma 2 is an open text model; it lacks the image and audio input the app needs and is not tuned for low-resource devices.

Within the open Gemma family, variants differ: Gemma 3n targets phones and other low-resource devices.

• AThe Flash-Lite model, built for low cost at high volume

  Incorrect

  Flash-Lite is Google's most cost-efficient model, optimized for high-volume, cost-sensitive, low-latency traffic; the firm has said quality matters more than cost or speed.

• BThe Gemini Live model, built for real-time voice interaction

  Incorrect

  Live models are optimized for live voice interactions; the firm needs written analyses of documents, and its analysts do not need instant replies.

• CThe Gemini Embedding model, built for semantic search

  Incorrect

  Embedding models turn content into vectors so related passages can be found; they do not write the multi-step analysis itself.

• DThe frontier Gemini model, built for deep multi-step work

  Correct

  Google's frontier Gemini model is described as excelling at deep work in complex real-world tasks such as finance and legal, with multi-step reasoning.

Gemini comes in tiers and specialized models. When the hardest reasoning is what the task demands and the business accepts more cost and latency for it, the frontier model is the fit; the other options are built for other jobs.

• AA Gemini embedding model, which turns text into vectors

  Correct

  Embedding models produce numerical vectors that capture meaning, so semantically similar content ends up close together and can be found with vector search.

• BVector Search, which finds the nearest stored vectors

  Incorrect

  Vector Search indexes and searches vectors very fast, but the embeddings have to be prepared first with a model; it is the next step, not this one.

• CA Gemini text model, prompted to tag articles with keywords

  Incorrect

  Keyword tags are text, not numerical representations of meaning, and two articles that say the same thing in different words can get different tags.

• DCloud Natural Language API, which extracts entities

  Incorrect

  Entity extraction lists the people, places and things a text names; it does not turn a whole article into a vector that captures its meaning.

Similarity search runs on embeddings: an embedding model turns content into vectors, then a service such as Vector Search finds the nearest ones.

• AA Gemini model served through the Gemini Live API

  Correct

  Gemini Live models are designed for real-time, bidirectional streaming with low-latency built-in audio for natural conversations.

• BA Gemini TTS model that turns replies into expressive speech

  Incorrect

  Text-to-speech models produce high-quality voice output, but only in one direction; they do not listen to the customer or hold a live conversation.

• CCloud Speech-to-Text feeding a Gemini text model

  Incorrect

  Speech-to-Text turns the customer's speech into text and the model replies in text; nothing in this pair speaks back in real time.

• DGemma 3n, an open model that accepts audio input

  Incorrect

  Gemma 3n takes audio as input but produces text output, so it cannot talk with customers on its own.

Real-time voice conversations need a model built for live, two-way streaming audio, which is what the Live API offers. Models that only listen or only speak cover half of the conversation.

• AGemma 2

  Incorrect

  Gemma 2 is a general open text model; it is not tuned to judge text and images against safety policies.

• BPaliGemma

  Incorrect

  PaliGemma is not tuned to judge content against safety policies.

• CShieldGemma 2

  Correct

  ShieldGemma 2 is an instruction-tuned Gemma model for evaluating the safety of text and images against defined policies, and as an open model it can be self-hosted.

• DGemma 3n

  Incorrect

  Gemma 3n is designed for efficient execution on low-resource devices; it is not tuned to judge content against safety policies.

The Gemma family includes specialized variants; knowing which one targets which job saves building a classifier from scratch.