Weights
Last updated June 11, 2026
What is Weights in simple terms?
In simple terms, weights are the dials inside an AI that get tuned as it learns. Each controls how much one signal matters to the next, and together they hold everything the model knows.
What is Weights?
Weights are the adjustable internal numbers of a neural network that determine how strongly each connection influences the next — the values a model tunes during training and the place where everything it has learned is ultimately stored.
A neural network is built from many simple units passing signals to one another, and every connection between them carries a weight — a number that says how strongly that signal should count. A large weight amplifies a connection's influence; a small or negative one dampens or flips it. On their own each weight is trivial, but a modern model has millions, billions, or more of them, and collectively they decide how inputs flow through the network to become outputs. When people say a model has "175 billion parameters," they're mostly talking about its weights. They are the substance of the model: not rules someone wrote, but numbers the system settled on.
Where weights come from is the whole story of training. A fresh model starts with random weights, so it behaves randomly. Training feeds it examples, measures how wrong it is, and uses gradient descent and backpropagation to nudge every weight a little in the direction that reduces the error — repeating this billions of times. Slowly the weights shift from random noise into a configuration that captures real patterns in the data. By the end, everything the model has "learned" — grammar, facts, styles, the feel of a cat's face — lives nowhere but in the specific values of those weights. There's no separate memory bank; the weights are the memory. This is why a trained model can be shared as a file: that file is essentially a giant list of its weights.
Understanding weights demystifies a lot about AI. It's why models are described by their size (more weights can capture more, at greater cost to train and run), why fine-tuning works (it gently adjusts existing weights for a new task instead of starting over), and why "open-weights" models matter (releasing the weights lets others run and adapt the model freely). It also explains a discomfort: because knowledge is spread across billions of opaque numbers with no human-readable labels, it's genuinely hard to point to where or why a model knows something, which is much of what makes AI difficult to interpret and to fully trust.
Real-world example of Weights
Think of a giant mixing desk in a recording studio, covered in thousands of sliders that each control how loud one instrument or effect is in the final mix. Set them randomly and you get noise; tune them just right and a clear song emerges. A neural network's weights are those sliders — except there are billions of them, and instead of a sound engineer adjusting them by ear, training does it automatically, nudging each slider over and over until the overall "mix" turns inputs into good answers. The finished song isn't written down anywhere as notes; it exists purely as the particular settings of all those sliders. That's exactly how a model's knowledge lives in its weights.
Related terms
Frequently asked questions about Weights
What is the difference between weights and parameters?
The terms are often used interchangeably, and weights are the main kind of parameter — so when a model is described as having billions of parameters, those are overwhelmingly its weights. Strictly, "parameters" is the umbrella term for all the adjustable numbers a model learns during training, which includes weights plus a few other values like biases. For everyday purposes, picturing parameters as the model's weights — the tunable dials that hold what it has learned — is accurate enough.
How do weights work?
Each weight sits on a connection between units in a neural network and scales the signal passing through it — a big weight makes that connection count for a lot, a small one makes it count for little. As data flows through the network, these weighted signals combine layer by layer to produce an output. The weights start random and are adjusted during training, via gradient descent and backpropagation, until they capture useful patterns. The model's behavior is entirely determined by their final values.
What are weights used for?
Weights are where a model stores everything it learns — its knowledge, skills, and style all live in their values, not in any written rules. They're what training adjusts, what fine-tuning gently re-tunes for a new task, and what gets saved when a trained model is shared as a file. The concept also underlies "open-weights" models, where the weights are released so anyone can run and adapt the model. In short, the weights are the model: change them and you change what the AI does.