What are model parameters? Explained with the dials on a mixing desk
Parameters are the billions of numbers an AI model tunes during training, and the 8B or 70B in its name counts them. Explained with the dials on a concert mixing desk.
- Author
Umair Mateen- Published
- Reading time
- 5 min
- Category
- AI
In one sentence: Parameters are the millions or billions of numbers inside an AI model that it adjusts during training, and the parameter count (the "8B" or "70B" in a model's name) tells you how big the model is, not how good it is.
The everyday picture
Picture the mixing desk at the back of a concert hall. It is a long table covered in small dials. Each dial changes the sound a little: a bit louder here, a bit softer there.
Before the show, a sound engineer runs a long rehearsal. They listen, hear what is off and nudge the dials. Over and over. By the end, the dials are set and the band sounds right.
An AI model is that desk. Each dial is a parameter. A big model is a desk with billions of dials. Training is the rehearsal that sets them.
Training nudges billions of parameters until the output sounds right. The settings are then saved, and every request you send runs through those fixed dials.
How it works, step by step
- The model starts with its dials at random (a brand new desk, nothing set). Its parameters are just numbers, and at first they produce nonsense.
- Training nudges every parameter (the rehearsal). The model sees huge amounts of example text, guesses what comes next, checks the real answer and shifts its parameters a tiny bit to do better. Google's machine learning glossary describes parameters as the weights and biases a model learns during training.
- People choose the setup, not the dial positions (the engineer decides how big the desk is and how long to rehearse). These choices are called hyperparameters. Google's glossary says people or a tuning service adjust them between training runs. Nobody sets the billions of parameters by hand.
- The settings are saved (the dials are locked in for the show). The finished set of numbers is the trained model. A file of "model weights" is these saved settings.
- Your request runs through the fixed dials (the live show). When you type a question, the model uses its saved parameters to produce an answer. Using a model does not change them. That is the difference between training and use, called inference.
Try it · Myth or fact
Model parameters: myth or fact?
Decide for yourself first, then tap each statement to check it.
0 of 4 checked
Where the picture stops
- A sound engineer turns one dial at a time, by ear. Training adjusts all the parameters at once, automatically, using maths rather than listening.
- Desk dials have labels like "bass" or "vocals". Parameters have no labels. No single number means "French" or "politeness"; skills come from many of them working together.
- A desk sounds better with more dials only up to a point. With models, size helps, but how much and how well the model was trained matters just as much (see below).
Why it matters for your business
- Bigger often scores higher, at a price. Meta's Llama 3 model card lists 8B and 70B versions, and the 70B version scores higher on its benchmarks, for example 82.0 vs 68.4 on the MMLU knowledge test for the instruction-tuned versions. Bigger models also need more computer memory and power to run, so they usually cost more per request and answer more slowly.
- Size is not the whole story. DeepMind's Chinchilla paper found many large models were "significantly undertrained". Their 70B model, trained on four times more data than their own 280B model, beat models with 175B to 530B parameters on many tasks.
- Hardware follows size. If you want to run a model on your own servers, the parameter count is the first clue to how much memory you need. Hugging Face's documentation gives an example: training a 4B-parameter model with its settings needed roughly 85GB of graphics card memory.
- Pick for the job. A smaller model may handle sorting emails or pulling fields from forms well and cheaply. Save the largest models for harder reasoning. Test both on your own tasks before deciding.
Try it · Try the dial
Turn the dial: what changes as a model gets bigger?
Drag the dial and watch what changes.
A small desk. Often runs on a single machine, answers fast and costs little. Good for focused jobs like sorting or extracting fields.
An example
Imagine a 15-person logistics company that wants AI to read delivery notes and fill in a spreadsheet. The team tries a large model and a small one on 50 real notes. Both get almost every field right. The small one answers faster and costs less per note, so they choose it, and keep the large model for the odd tricky case.
Frequently asked questions
What does 7B or 70B mean in an AI model's name?
It is the parameter count: B stands for billion. A 70B model has about 70 billion parameters, the adjustable numbers it learned in training.
Are parameters the same as tokens?
No. Parameters are the numbers inside the model. Tokens are the small pieces of text the model reads and writes. Training data is measured in tokens; model size is measured in parameters.
Are more parameters always better?
No. More parameters can capture more patterns, but a smaller model trained well on more data can beat a larger one. Cost, speed and your actual task matter too.
What is the difference between parameters and hyperparameters?
Parameters are learned by the model during training. Hyperparameters are settings people choose before or between training runs, such as how big the model is and how long it trains.
Related explainers
The DevDrift helps teams choose and run the right-sized AI model for each job: see how.
Sources
Could you explain this to a colleague now?
Mark it and posts that mention it will show a tick. Saved in this browser only.