What are AI parameters and weights? The numbers inside a model

Think of adjustable settings inside a model

Imagine a simple calculator for packing boxes. It uses one number to describe the space needed for each item. Changing that number changes its estimate.

A learned model also uses numerical settings to turn input into a result. The values learned during training are called parameters. Weights are a common kind of parameter.

The packing calculator is a teaching example. It does not show the full structure of a language model, which uses many connected calculations.

A weight changes how strongly an input affects a result

In a small model, a weight can make one input have a stronger or weaker effect. Other learned values can shift the starting point of a calculation.

A simple picture is a volume dial. Turn it, and one signal has more or less influence. The analogy has limits: a model does not have a single dial labelled “be truthful” or “know geography”.

The model’s behaviour comes from many values working together. Changing one number is not a reliable way to add a specific fact.

What does 8B mean?

B stands for billion. An 8B model label usually means roughly eight billion parameters. The precise count can differ from the rounded name, so check the model’s official description.

The number is not a count of words, stored facts, or correct answers. It also does not mean an 8 GB download. File size depends on how the values are stored and what else the file contains.

Later, the lesson on quantization will explain one way those stored values can take up less space. For now, keep parameter count and file size separate.

Does a larger model always give better answers?

No. The way a model is trained matters too. In the 2022 Chinchilla research, a smaller model trained with more data outperformed several larger models on the tests in that paper.

This was a specific research comparison, not a promise that every small model beats every large one. Its useful lesson is that parameter count alone cannot tell you which model will do your job better.

When choosing a model for shop notices, compare how well it follows your instructions and preserves your facts. A name with a bigger number is not enough evidence.

Are weights the same as your chat settings?

No. Asking for a short answer changes your request. Selecting an output-length limit changes how the app runs that request. Neither action is the same as training the model’s learned values.

In ordinary use, you load an existing model and ask it to respond. You do not have to adjust its billions of values by hand. Further training is a separate process.

An app may save your preferences or previous messages. That can affect later requests without changing the saved model weights.

Try it: read a model description carefully

Find an official model description, without downloading anything. Look for its parameter count and the size of a particular downloadable file. Write them on separate lines.

If the page offers several files for the same model, compare their sizes. Do not assume the largest one is best for your computer. The file instructions and your hardware both matter.

If you cannot find a clear count, write “not confirmed”. It is better to leave a blank than to turn a guess into a fact.

• Parameter count: how many learned values.
• File size: space occupied by the chosen file.
• Task quality: results on questions you can check.