Google publishes several families of foundation models and sells access to them on different terms, and the Generative AI Leader exam names four of them. Gemini is the proprietary multimodal family that Google runs as a managed service. Gemma is a family of open models whose weights Google publishes for anyone to download. Imagen generates and edits still images, and Veo generates video. An exam question describes a business situation and expects one of those four names, so a candidate needs to know what separates them.
The sections below take the four models in turn, starting with Gemini and the choice of size inside that family, then Gemma and the decision between an open model and a managed one, then Imagen and Veo. The last section sets out the signals that tell a candidate which of the four a question is asking for.
Gemini
Gemini is Google's flagship family, and it is multimodal by design. One model reads and reasons over text, images, audio and video in the same request. Gemini is also what sits behind Google's own assistants, meaning the Gemini app, Gemini for Google Workspace and Gemini Enterprise.
The family comes in several sizes, and the size is where the commercial decision sits. A larger Gemini model reasons better over hard, open ended work. A smaller one costs far less and answers faster, and on a narrow task it is frequently just as good. When a question mentions high volume, tight latency or cost pressure, it is usually asking for a smaller size within Gemini.
Gemma
Gemma is a family of open models, and the difference from Gemini is the licence and the weights. Google publishes the weights, so an organisation can download a Gemma model, run it on its own infrastructure or on another cloud, inspect it and adapt it.
Three considerations push an organisation towards Gemma.
- Sovereignty, where data may not leave a jurisdiction or a network.
- Cost at scale, where a per token bill outgrows the cost of the hardware.
- Reach, where inference has to happen somewhere a managed API cannot go, such as a restricted network or a device.
An organisation choosing Gemma takes on the hosting, the scaling, the updating and the security of the model, all of which a managed service had been doing on its behalf.
The line between open and managed is the one the exam returns to. When a question describes a regulated environment where data cannot leave, or hardware the organisation controls, it is pointing at Gemma.
Imagen
Imagen is Google's image model. It produces a picture from a text description, edits a picture that already exists and generates variations on an image a team has approved.
Editing is the part candidates forget. When a team already holds one photograph and needs twenty versions of it for twenty markets, the answer is still Imagen.
The business cases are marketing and creative production, product imagery and any work a team currently commissions from a photographer or licenses as stock.
Veo
Veo is Google's video model. It generates video from a text description or from a still image, at a higher fidelity and a longer duration than earlier generations of the technology.
The business cases are short form marketing, product demonstration and prototyping. An organisation can show a stakeholder a short video of a concept before it commissions a production shoot, and the decision to spend on the shoot then rests on something the stakeholder has seen.
Matching a described situation to a model
Each of the four models carries a signal that a question will use. Reasoning, summarising, extraction, question answering or anything mixing modalities in a single request points at Gemini. A need for the weights themselves, for self hosting or for a restricted environment points at Gemma. Still images are Imagen, and moving images are Veo.
Where two of the four seem to fit, the question usually states the deciding factor, and that factor is commercial more often than it is technical. It is the cost, the latency, the place where inference may happen or the organisation's own need to hold the model.