What We Mean by a Model
A model is the finished, trained thing. What people really mean by the model, and by language model.
A quick note on how we teach here. You won't find equations or numbers in this series, and that is on purpose. The goal is understanding, not arithmetic, so every idea is explained in plain words and simple pictures. When you want the math later, the rest of Internals Decoded is ready for you.
Welcome back. Last time we saw that training is like studying, and using the machine afterward, which people call inference, is like taking the exam. The training phase is loud and busy. Knobs get adjusted, mistakes get made, and the whole machine slowly learns the patterns hidden in the examples. But eventually that phase ends. The learning stops. What are you left with? Something quiet and ready. Something you can actually use. That something is what we are going to talk about today.
A model is the finished, trained thing. All that studying and adjusting leaves behind a settled arrangement of knobs that captures the patterns it saw. That arrangement, no longer changing, is the model.
The chef, not the cooking school
Imagine a chef who has spent years in a busy kitchen. She has chopped vegetables every way you can imagine. She has tasted sauces, burned a few dishes, and learned exactly how much heat each pan needs. All that practice is behind her now. You cannot point to the training. It is over. But you can point to the chef. She carries all of it inside her. When you hand her a basket of ingredients, she can turn them into a meal without thinking twice. The model is like that chef. It is not the years of schooling. It is the person who emerged from the schooling, ready to cook on demand.
The same idea applies to the trained machine we have been building. The model is the arrangement of knobs that resulted from training. It is the thing you can hold onto, copy, and hand to someone else. They do not need to see the examples or watch the mistakes. They just get the finished result.
In practice, a model is saved as a file, or a set of files, holding the setting of every knob, together with a short description of how the knobs are arranged, such as how many layers there are. When people say they downloaded a model, that is what they downloaded, as Hugging Face's guide to loading models shows.
What the model really is
A model is not a magic being. It is not a dusty encyclopedia full of stored answers. It is the outcome of training. Think of a huge, patient arrangement of settings. Every knob we talked about earlier is now set to a particular value. None of those values were written by a person. They were discovered by the machine as it stumbled through example after example, making guesses and correcting itself.
When you give the model an input, it does not go look up a prewritten reply. It passes your input through all those settled knobs, layer after layer, just like we explored in the pieces about tiny decision makers and deep stacks. The knobs work together to turn your input into a sensible next step. That might be the next word in a sentence, or a label for a picture, or whatever it was trained to produce. The model is the whole set of knobs, frozen in time, ready to do that work.
So when you hear people say "the model," they mean this trained result. They mean the chef, not the cooking school. They mean the quiet, settled thing that can take an input and produce an output without any more learning.
A language model
Now let us connect this to language. You have seen how a machine can learn to guess what word comes next in a sentence, back in Guessing What Comes Next. When a model is trained on a huge amount of language, doing exactly that kind of guessing, the finished result is called a language model. It is a model whose knobs have been shaped by the patterns of words and sentences. It does not understand language the way you do, but it has captured the rhythms and regularities. It can take a few words you give it and continue the thought in a way that feels natural.
If the model is very large, with a huge number of knobs and many layers, people call it a large language model. You will usually see it shortened to LLM. That is the phrase you hear everywhere today. It is just a language model that had more room to learn and soaked in a wider sea of text. The idea is the same. A settled arrangement of knobs, trained to guess the next word, ready to use.
One more thing. The chat assistants you use went through a second round of training, on examples of helpful answers and on people's ratings of which replies were better, as that same piece described. That is why they answer your questions instead of just carrying on your sentence. Underneath, each one is still a language model, guessing one word at a time.
Where this is heading
A model is the finished, trained thing. It is the quiet arrangement of knobs that you can hand to a friend and say, "Here, try it." It is the chef, not the cooking school.
But there is still one piece missing. Before a language model can work with the words you type, it has to turn those words into something it can handle. It has to chop the language into small, manageable pieces. That chopping has a name you keep hearing, and it is the very next step.
Test yourself
Which of these is the model: the huge pile of examples, the months of practice, or the finished set of knob settings? Why?
Answer: The finished set of knob settings is the model. The pile of examples is what the machine practiced on, and the months of practice are the training, the cooking school in our picture. Both are over once training ends. What is left is the arrangement of knobs they produced, saved as files you can copy and share. Like the chef who walks out of cooking school, the model carries what practice built, and it can turn new ingredients, or new questions, into answers without needing the examples or the practice again.
Next: Breaking Language Into Pieces. It shows how a language model chops your words into small standard pieces, called tokens, before it does anything else.
