The words behind what you have been making

The first two modules had you making images and then looking inside the machine that makes them. Some of their words, such as adapter, seed and guidance, turn up again here, alongside many you have not met yet. Most conversations about AI go wrong on vocabulary before they go wrong on anything else, because the same word means one thing to a researcher, another to a vendor and a third to the client in the room. Nobody stops to check. This module collects the words in groups, arranged by where you will hear them, so that you can follow a conversation instead of looking terms up one at a time. It starts here, with the words for the model itself, because they tell you what the thing is, who made it, where it runs and what you are allowed to do with it.

Start with the object. A model is its weights, the enormous set of numbers that training produced, and everything around them is software. A checkpoint is a saved state of those numbers, taken at some point during training, and people use the word loosely to mean any model file. On Hugging Face, where most open models are published, a model is a page with a card that describes it and a list of files, and the large files in formats such as .safetensors are the weights. FLUX.1 [schnell], an open image model from Black Forest Labs, has 12 billion parameters, which is another way of saying 12 billion weights. The figure looks impressive in a sales deck and tells you less about quality than you would expect.

A base model is the general one, as its maker released it, before anybody adapted it. Most of the products you will be offered are built on a base model that belongs to somebody else, with an interface, some instructions and a price added, and people call these wrappers. That is often useful, but it means the product's name tells you less than the name of the model underneath. Versions matter just as much. Google's documentation for its Gemini models separates stable names, which usually do not change, from a latest alias that is swapped for each new release, and it deprecates preview models with at least two weeks' notice. Pinning a version means calling the model by its exact stable name, so the look you approved in March is the look you still get in June.

Then there is where the model runs. A hosted model runs on its maker's computers and you reach it through an API, a web address your software sends requests to, paying per call; you never hold the file, and the maker can change or retire it. An open-weights model is one you can download and run yourself, on your own graphics card or a rented one, and the file you downloaded will not change underneath you. Open weights is a different claim from open source. The Open Source Initiative published a definition for AI in 2024 that asks for the weights, the complete code used to train and run the model, and enough detail about the training data for a skilled person to build a substantially equivalent system. FLUX.1 [schnell] ships its weights under the Apache 2.0 licence, but its card does not describe the data it learned from.

Two more words explain most of the price differences you will see. A distilled model is a smaller or faster model trained to imitate a larger one. FLUX.1 [schnell] was made this way, by what its makers call latent adversarial diffusion distillation, and it produces an image in one to four steps where an undistilled image model typically needs twenty or more. A quantised model has had its numbers stored at lower precision, 8 or 4 bits each instead of 16, so that it fits on a smaller graphics card or a laptop, at some cost in quality. Both of these are trades. Speed and price are bought with a little quality, and the loss tends to show first on the hardest jobs.

In a working meeting you need four questions, and each one uses this vocabulary. Ask whether the model is open or hosted, because that decides who can change it. Ask if the version can be pinned. Ask whether it has been distilled or quantised, which usually explains why it is cheap, and ask who made the model underneath, because a great many products are one company's interface on another company's model. A designer who can ask those four in plain words, and understand the answers, will spot most of the traps before any budget is spent.