Plain-English explanation
A generative model doesn't store images, text, or rules the way a database does. It stores weights: enormous grids of numbers that determine how the model transforms an input (a prompt, a noisy image, some tokens) into an output. Training is the process of nudging those numbers, over millions of examples, until the model produces useful results. Once training is done, the weights are the entire model, everything the model "knows" is encoded in that set of numbers.
It helps to separate the concept from the file. "Model weights" describes what's inside; "checkpoint" describes the file you download and load. You'll often hear the terms used loosely and interchangeably, but weights is the underlying data and a checkpoint is one particular saved packaging of it.
How it works
Structurally, weights are organized into the layers of a neural network: for a diffusion image model, that includes the weights of the UNet or transformer backbone, the text encoder, and the VAE; for a language model, the weights of each transformer layer's attention and feed-forward blocks. Each weight is a floating-point number, and a modern model can have anywhere from a few hundred million to well over a hundred billion of them, which is why full model files run from hundreds of megabytes to tens of gigabytes.
The numerical precision those weights are stored at (32-bit, 16-bit, 8-bit, and so on) directly affects file size and hardware requirements; reducing that precision is called quantization. Weights can also be modified after initial training through further fine-tuning, or adjusted at load time by applying a small add-on like a LoRA without permanently changing the original weights at all.
Adult-AI use
Most users never touch weights directly; they interact with a checkpoint file, a slider, or an app. But the concept matters because it's what people acquire, adapt, or protect when they talk about "the model." Whether an AI companion app can be run entirely on your own hardware, whether a character can be customized with a LoRA, and whether a local image generator can produce NSFW content at all, all come down to what's encoded in that model's weights and whether you have access to them at all (see open-weight model).
How it differs from a checkpoint
Think of weights as the content and a checkpoint as the container. Weights are the abstract, general concept of a model's learned parameters. A checkpoint is a specific saved file containing a specific set of weights at a specific point: the base model release, a fine-tuned variant, or a merge. You never load "weights" directly; you load a checkpoint file that contains them.
Limitations and misconceptions
- Weights alone aren't a working model. You also need the code (architecture definition) that knows how to interpret those numbers; the same weights are useless without matching model code.
- More weights isn't automatically better. Parameter count correlates loosely with capability, but training data quality and fine-tuning matter enormously too.
- Weights can be frozen or trainable at different stages. During inference, weights are fixed; during training or fine-tuning, they're actively being updated.