Google's Foundation Models: Gemini, Gemma, Imagen, and Veo Explained

Google's foundation models are large models trained on broad datasets that provide general-purpose generative capability, and the four the exam names are Gemini, Gemma, Imagen, and Veo. A foundation model is a starting point: rather than building a model for one narrow task, a business takes a foundation model and applies it, through prompting, grounding in company data, or tuning, to many tasks.

The four models divide the landscape along two lines. The first is modality, the type of content a model understands or produces: Gemini understands text, images, audio, video, and code; Imagen produces images; Veo produces video. The second is openness: Gemini, Imagen, and Veo are Google-managed models whose weights are never released, while Gemma is an open model whose weights can be downloaded, modified, and run wherever the organization chooses.

All four are available to builders through Model Garden on Agent Platform, and Gemini also powers Google's end-user applications such as the Gemini app and Gemini for Google Workspace.

Gemini is Google's flagship foundation model family, and its defining strength is that it is natively multimodal: it was built from the start to reason across text, images, audio, video, and code together in one request, rather than handling one type of content at a time or passing each type to a separate system. Typical uses follow from that: summarizing and answering questions over long documents, analysing charts, screenshots, and scanned forms, drafting and reviewing code, and powering conversational assistants.

Its second strength is a long context window: one request can hold a very large input, such as a lengthy contract, a full codebase, or an hour-long recording, together. That matters because the alternative, splitting a long document into chunks that are processed separately, loses the links between distant parts: an obligation in an early section that a definition in a late appendix narrows, or a remark early in a meeting that contradicts a decision at the end. A single long-context request keeps the whole input in view, so the model can connect those parts.

Several Gemini capabilities sound similar but solve different problems:

Gemini is also the reasoning engine inside agents. In practice that means the model interprets the request and any mixed inputs, decides which tool to call and with what parameters, checks the result it gets back, and re-plans if the result is incomplete or wrong, repeating until the task is done. None of this requires open weights: tool use is a capability of the managed model, so needing an agent is not a reason to choose Gemma.

Gemini comes in more than one size and setting, and these are easy to confuse with Gemma:

• Lightweight managed tiers such as Flash-Lite cost less per request, but they are still served by Google and their weights are not released.
• Gemini Nano is a compact Gemini built into Android devices and the Chrome browser. It runs on-device with no network connection, but it is delivered by the platform, not a model an organization downloads, tunes, and installs on its own hardware, and it is not meant for peak reasoning over long inputs.
• Gemini on Google Distributed Cloud, including an air-gapped option, brings Google-supplied, certified hardware and Google-operated software into a customer's data center, so prompts and outputs stay inside the customer's perimeter. Google deploys and operates Gemini there, and the weights are never released for the customer to download, inspect, modify, or redeploy elsewhere. It is not a way to run Gemini on servers the organization already owns; that requirement still points to Gemma.

The rule these share: small size, low cost, or running on-premises is not the same as open weights. Only Gemma lets an organization possess and modify the model itself.

A concrete example: an insurance company processing claims receives photos of vehicle damage, a written incident description, and a phone-call recording. Because Gemini is multimodal, one model can review all three, extract the relevant facts, flag inconsistencies between the photos and the description, and draft a structured claim summary for the adjuster. Before multimodal foundation models, that workflow would have required separate systems for images, text, and audio, plus custom work to combine their outputs.

Gemma is Google's family of lightweight open models, built from the same research and technology behind Gemini. Open means the model weights are published: an organization can download Gemma, run it on its own infrastructure, in a chosen cloud environment, or on capable local devices, and fine-tune it on its own data. Gemma generates text; most current Gemma models also accept image input, and some variants (such as Gemma 3n) also take audio, and the family includes specialised variants built for particular jobs, such as on-device use or particular domains.

Gemma's strengths follow from that openness. It gives organizations control over where the model runs, how data flows, and how the model is customized, and its lightweight sizes need far less compute than flagship models, so it can run on modest hardware, at the edge, or on a device with limited connectivity.

Tuning shows the difference most clearly. Tuning a managed Gemini model produces a tuned model that the platform serves to you through an endpoint; you use it, but you do not hold it. Tuning Gemma produces a set of weights the organization keeps: it can version them, audit them, and deploy them anywhere, including on its own servers or devices. Grounding and few-shot prompting are different again: they shape a single answer with retrieved data or examples and change no weights at all, for either model.

The business use cases are situations where control outweighs raw capability. A hospital network that must keep patient data on servers it runs can host a fine-tuned Gemma model there. A manufacturer can deploy a small Gemma model on equipment near the factory floor where connectivity is limited. A product team can fine-tune Gemma into a narrow specialist it owns and ships inside its own application. Cost efficiency is a genuine Gemma strength in one particular shape of problem: a single narrow, repetitive task at very high volume, such as classifying or tagging short messages, where a small model tuned for that one job performs well enough and costs far less per request than a large general model. The trade-off: Gemma offers deployment freedom and ownership of the weights, but the organization takes on model operations (hosting, scaling, patching, monitoring); Gemini offers the strongest capability with no operational burden.

Imagen is Google's text-to-image foundation model: you describe an image in natural language and Imagen generates it, with strengths in photorealism and faithfulness to the prompt. Imagen also edits existing images from an instruction, keeping the parts you did not ask to change. Because the interface is plain language, non-designers can produce usable visuals, and design teams can produce far more options in far less time.

Generation and editing are different jobs, and the editing modes are worth knowing by name:

When the brief is to keep an existing photograph and change only part of it, that is editing; generating a fresh scene would lose the original.

The business use cases center on visual content at scale. Marketing teams generate campaign imagery and seasonal variations without scheduling photo shoots. Retailers produce product visuals in different settings. Product and UX teams turn written concepts into visual mockups during early exploration. A furniture retailer preparing a holiday campaign, for example, can generate its sofa in a dozen styled living rooms, select the strongest directions, and reserve the physical shoot for the one hero image. As with all generative media, businesses should pair Imagen with responsible-use review before publishing.

A currency note: Google has deprecated the Imagen models as a whole, generation as well as editing, on both the Gemini Developer API and the Agent Platform (formerly Vertex AI) Gemini API, and directs new image work to its Gemini image models, which generate and edit images. The exam guide still names Imagen as Google's image generation and editing model, so treat Imagen as the image answer for the exam, but do not treat "Gemini cannot make images" as a rule.

Veo is Google's video generation foundation model: from a text description, or from a starting image, it produces short high-definition video clips, with control over cinematic qualities like camera movement and visual style. Current Veo models (Veo 3 and later) also generate synchronized audio with the picture: dialogue, sound effects, and ambient sound, so a clip can come out with its soundtrack already matched to the action.

The rule to hold onto: Veo creates new footage; Gemini understands existing footage. Summarizing a recorded webinar, finding the moment a defect appears in inspection video, or answering questions about a clip is video input, which is Gemini's multimodal understanding. Producing a clip that did not exist before is video output, which is Veo. Likewise, Imagen produces still images and Veo produces moving video.

The business use cases mirror where video budgets go today. Marketing teams generate short-form video for ads and social channels without a production crew. Creative teams storyboard and pre-visualize concepts in motion before committing to a shoot. Training and communications teams turn scripts into illustrative footage. E-commerce teams animate a product photo to show the item in use.

A concrete example: a travel company wants ten short video variations promoting different destinations for a social campaign. Filming on location in ten countries is out of budget; with Veo, the team generates destination clips, complete with ambient sound, from descriptions, tests which perform best with audiences, and invests real production money only in the winners.

The factor that usually decides among Gemini, Imagen, and Veo is the output the task requires. Generating or editing still images points to Imagen. Generating video, with or without its own audio, points to Veo. Understanding, reasoning, conversation, code, or analysis of mixed inputs, including existing images and video, points to Gemini. This is a practical heuristic for reading scenarios, not a decision order Google publishes.

Openness matters mainly between Gemini and Gemma for language work. Managed Gemini is the default: strongest capability, no infrastructure to operate, enterprise controls provided by the platform. Choose Gemma when a requirement calls for what only open weights give: possessing and modifying the model, running it offline or on an edge device you deploy to, keeping it on infrastructure you run, or owning tuned weights you can version and redeploy. Tuning never adds an output modality: fine-tuning Gemma cannot make it produce video, so a video requirement still points to Veo.

Cost needs care in both directions. When the task is narrow, repetitive, and very high volume, such as tagging or routing huge numbers of short messages, fine-tuning a small Gemma model into a specialist for that one job is a sound, cost-efficient choice: a small tuned model is enough, and its per-request cost is far lower. Cost alone is not a reason to self-host Gemma when the work needs strong general reasoning across varied requests and the team cannot staff model operations (serving, scaling, security patching, monitoring); there, a lightweight managed Gemini tier may already meet the budget. If no requirement such as data locality, offline or edge operation, owning tuned weights, or a narrow high-volume specialist task calls for Gemma, managed Gemini is the better fit.

Beyond modality and openness, the usual selection factors still apply: context window (how much material the model can consider at once), cost and performance for the expected volume, security and data-handling requirements, availability and reliability expectations, and how much customization the use case demands. The models are complements, not competitors: a single campaign might use Gemini to develop messaging, Imagen to create the imagery, and Veo to produce the video, and a business choosing Google Cloud gets all four through one platform.

A consumer-electronics company launching a smart-home device shows how the four models divide the work. The launch needs market analysis, a support experience, launch visuals, and promotional video.

The strategy team uses Gemini to analyze competitor materials, customer reviews of the previous model, and internal research documents, then drafts positioning and messaging from that grounded picture. The support organization builds an assistant, powered by Gemini as its reasoning engine, that answers setup questions by drawing on the product manual and troubleshooting guides.

The creative team uses Imagen to generate launch imagery, the device styled in modern kitchens, living rooms, and offices, iterating through directions before commissioning final photography, and uses Veo to produce short teaser clips showing the device in daily-life scenes for social channels. Meanwhile, the engineering team has a constraint: an on-device voice feature must work without sending audio to the cloud. They fine-tune Gemma into a compact model that runs on the device itself, keeping user audio local.

One launch, four models, each chosen by modality and control needs: Gemini for reasoning and conversation, Imagen for images, Veo for video, and Gemma where open weights and on-device deployment were non-negotiable. That mapping, use case to model, is precisely the judgment the exam tests.