The models layer is where the actual AI models live, and it is one of the most heavily tested areas on the Generative AI Leader exam. When I work through this section of the material, I want a clean mental model of three things: what foundation models are and why they generalize the way they do, how specialized or fine-tuned models relate to them, and what parameters actually represent inside a model. Those three pieces show up over and over again in scenario questions.
This article walks through all three, faithful to how the concepts are framed in my Generative AI Leader course.
The AI ecosystem is a stack of layers, each building on the one below it. From the bottom up: infrastructure, models, platforms, agents, applications. As you move up, you get greater abstraction and you get closer to the end user.
The models layer sits directly on top of infrastructure. This is where pre-trained Gen AI models like OpenAI's GPT series, Claude, or Llama actually live. Infrastructure provides the massive computational resources needed to train these models. Once trained, the models become the foundation that platforms, agents, and applications are built on top of.
One distinction the exam likes to test: this layer is about the underlying models themselves, not the applications or user interfaces built on top of them. GPT-5 is in this layer. ChatGPT, the product you interact with, is not. ChatGPT is an application layer concern. Same pattern with Gemini versus the Gemini app, or Claude versus claude.ai.
The examples given in the course span the major providers. From Anthropic, Claude Opus and Claude Sonnet. From Google, Gemini, Imagen, and Veo. From OpenAI, GPT and Sora. These are the underlying models, each trained on massive datasets and capable of a wide range of tasks before any specialization gets layered on.
Foundation models are large, general-purpose systems trained on diverse data that can handle many different tasks. The mental picture I keep in my head is a flow diagram with three pieces.
On the left, massive, diverse data flows in: books, articles, websites, code, images. In the middle, that data trains a foundation model, which is a deep neural network that learns patterns across all of it simultaneously. On the right, what comes out is the ability to perform a variety of tasks: writing, coding, translation, image generation, and more.
The reason a single model can do all of that is precisely because of what went into training it. When a model has seen enough diverse data, it develops internal representations that generalize across domains rather than being locked to one narrow task. This is the takeaway worth memorizing for the Generative AI Leader exam: diverse training data is what explains the adaptability of foundation models, not the architecture alone. If you see a question asking why a foundation model can handle so many tasks, the answer points back to the training data, not to anything special about the network shape.
Fine-tuning a foundation model means taking a general model and continuing to train it on more specific data so it performs better on a particular task. That task could be writing code, generating images, analyzing scientific data, or anything else where you have domain-specific data on hand.
The reason to fine-tune is straightforward. If you are working with domain-specific data, a fine-tuned model will outperform a general one. You are honing the model to better match the new task and the particular characteristics of your data.
One nuance worth flagging: there is no bright line between a foundation model and a specialized one. Fine-tuning is a spectrum. A heavily fine-tuned model might look quite different from its base, but it started from the same foundation. The course frames this as a continuum rather than a binary, and the exam reflects that framing.
The example I find clearest is the image classifier. You start with a generic image classifier trained on everyday images: handwritten digits, x-rays, cats. You then fine-tune it to your domain, say astronomical images of stars and galaxies. The result is better predictions on that domain than you would get from the general model alone. The weights from pre-training are not thrown away, they are adjusted. That is what makes fine-tuning efficient compared to training from scratch.
On Google Cloud, fine-tuning is available through Vertex AI for both Gemini models and open models like Llama. If you see a question about adapting a model to a domain on GCP, Vertex AI is the platform-level answer.
The other concept in the models layer that comes up constantly is parameters. Modern Gen AI models are usually measured and compared by their parameter counts, which typically run into the billions or trillions. Generally, more parameters means more sophistication, though the relationship is not purely linear.
To understand what a parameter actually is, it helps to picture a neural network. The network is made up of layers of nodes connected by lines, and those connections are where the parameters live. Each connection has a weight, which is a numerical value that determines how strongly the signal from one node influences the next. During training, these weights get adjusted billions of times until the model's outputs match the expected results as closely as possible. The output of the network is the end result of all those weighted signals flowing through.
When a model has more parameters, it has more weights, more connections, more layers, and more capacity to represent complex patterns in data. A model with a trillion parameters has far more room to encode nuanced relationships in language or images than one with a billion. That is the underlying reason parameter count correlates with capability. It also means more compute is required to train and run the larger model, which is a real tradeoff that comes up in cost and performance scenarios.
On Google Cloud, you can access models at different parameter scales through Vertex AI Model Garden. Model Garden lets you choose based on your performance and cost requirements rather than being forced into a single parameter tier.
The three concepts in this article fit together as one coherent picture of the models layer. Foundation models are general-purpose systems whose adaptability comes from diverse training data. Specialized models are foundation models that have been fine-tuned to a specific domain, with no hard line between the two categories. Parameters are the weights inside the model, and parameter count correlates with capability and with compute cost.
If you can recognize which of those three the exam is asking about and answer in the framing the course uses, this section is worth easy points.
My Generative AI Leader course covers foundation models, fine-tuning, and parameters in depth alongside the rest of the foundational material you need for the exam.
