Parameters
Also called Weights
The billions of numbers a model tunes during training — its entire learned knowledge, frozen into a giant grid.
Think of it like
The settings on an impossibly complex mixing board, dialed in once and then locked.
Example
“A 70B model” means 70 billion of these numbers doing the work.
How it actually works
Each parameter is a weight on a connection inside the network. Training nudges all of them, gradually, until the model’s guesses line up with real text. At inference they never change — the model is a fixed function, and the parameters are the function. More parameters usually means more capacity, but also more cost to run.
For product teams
A rough proxy for capability and cost — bigger models know more but are slower and pricier.
For engineers
Learned weights/biases updated by gradient descent in training, static at inference.
Related
- Fine-tuning — Adjusted a little during fine-tuning.
- Neural network — The raw material a neural network is made of.
- LLM — The “large” in large language model.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome