What are AI model weights?

Model weights are numbers learned during training that influence a model’s calculations. During use, those numbers help turn an input into a prediction or generated output.

The useful distinction is between the learned numbers, the software that uses them and the terms governing their use. Obtaining one does not settle the other two.

Think of adjustable controls, within limits

Imagine a panel of adjustable controls. Changing a control changes how a machine responds to an input. Training adjusts numerical values; using the trained model applies the resulting settings.

That analogy explains influence, not understanding. It does not mean each weight controls one recognisable fact or skill, or that a person can read a weight file like a handbook. The weights act together within the model’s structure.

Google’s machine-learning glossary describes training as determining weights and inference as using them to make predictions. Its linear-model example shows an input value multiplied by a learned weight.[B4-1]

A small worked example

Invented numbers for arithmetic only; no model was trained or tested.

Suppose a simple model uses two inputs and calculates:

output = (first input × first weight) + (second input × second weight) + bias

The bias is another adjustable parameter: a number added to the calculation.

Part	Illustrative value	Contribution
First input and weight	4 and 0.5	4 × 0.5 = 2
Second input and weight	2 and −1	2 × −1 = −2
Bias	3	+3
Output		2 − 2 + 3 = 3

Keep the inputs and bias fixed, but change the first weight from 0.5 to 1. The output becomes (4 × 1) + (2 × −1) + 3 = 5.

The point is that a changed weight changes this calculation. It does not show a better answer, a probability or a business result. A language model involves many connected calculations; this small example does not reproduce one.

What do “parameters” and “quantisation” mean here?

A parameter is an adjustable value in the model. Weights and biases are examples. A parameter count describes how many such values there are; it is not an evaluation of the model’s work.

Quantisation represents numerical values with reduced precision, often using fewer bits per value. It changes their representation, so record which variant you are evaluating. Hugging Face’s quantisation documentation explains numerical precision and representation; it does not establish the behaviour of your selected files.[B4-2]

Next reading: local AI memory-budget worksheet.

Downloaded weights are one part of the arrangement

Keep three questions separate:

• What did you obtain? Record the publisher, exact files, revision and representation.
• What may you do with it? Check the applicable licence and companion terms for the planned activity.
• How will you use it? Identify the runtime, surrounding application and operating owner before evaluating the workflow.

Under the Open Source Initiative’s Open Source AI Definition 1.0, the preferred form for modification includes data information, code and parameters. Downloadable weights alone do not establish that a system meets that framework.[B4-3]

For the next step, return to open weights, closed weights and rights, record permissions in the exact-weight licence worksheet, or examine hosted access versus operating weights. Use Running AI locally for the wider workload decision.

Sources and boundaries

• [B4-1] Google’s machine-learning glossary, weight entry: learned weights, inference and the linear-model calculation. Our control-panel analogy and numbers are original teaching illustrations.
• [B4-2] Hugging Face quantisation concepts: reduced precision and numerical representation. No performance comparison is asserted here.
• [B4-3] Open Source AI Definition 1.0: OSI’s preferred form for modification, not a licence for a particular checkpoint.

This explainer does not clear files, permissions, hardware or a deployment.