What Are Weights (in AI)?

What it is: Weights are the numerical values inside an AI model that determine how strongly each input feature contributes to the model’s output. They are the model’s “learned knowledge.”
Who it’s for: Anyone trying to understand what’s actually inside a large language model
Best if: You’ve heard people talk about “open-weights” models and want to know what’s being released
Skip if: You already train neural networks for a living

What Are Weights (in AI)?

Weights are the dials inside a neural network. A modern AI model is a stack of mathematical layers, and each connection between those layers is multiplied by a number called a weight. Training is the process of nudging billions of those numbers up and down until the model produces useful output. Once training stops, the final set of numbers — the weights — is the model. Everything the system “knows” about language, code, images, or chemistry lives in that file of numbers.

A useful analogy: imagine a giant mixing console with a billion sliders. Before training, every slider is set randomly and the console produces noise. Training shows the console examples — “when the input looks like this, the output should sound like that” — and after each example, an algorithm called gradient descent nudges every slider a tiny bit. After trillions of examples, the sliders settle into a configuration that turns inputs into intelligible outputs. The slider positions are the weights.

This is why people talk about a model’s parameter count in the billions. A weight is a parameter; “70B parameters” means roughly 70 billion individual numbers. Together they form the neural network that produces a response when you type a prompt.

Why It Matters

Weights are why “open-weights” is a real distinction in the AI industry. When Meta releases Llama or Mistral releases a model, they publish the weight file. Anyone can download those numbers, load them into a runtime, and get the same behavior the original lab gets. In contrast, closed models like Claude or GPT-5 keep their weights private — you only ever interact with them through an API. So when you hear “open source AI,” what is usually being shared is the trained weights, not the training data or the training code.

Weights also explain why training is so expensive and inference is so commodity. Training has to nudge those billions of numbers across trillions of examples — that takes months and tens of millions of dollars in compute. Once trained, the same weights can serve millions of users on a single GPU because inference just multiplies the input by the weights and reads the output.

How It Works

Each weight sits between two artificial neurons. When information flows through the network, every signal is multiplied by the weight on its path. Large weights amplify a signal, small weights mute it, negative weights flip its sign. Stack a few hundred billion of these multiplications together across dozens of layers, and you get the rich behavior we call “reasoning.”

During training, the model makes a prediction, compares it to the right answer, and computes how wrong it was. An optimization algorithm — usually a variant of gradient descent — then walks backward through the network and adjusts each weight in the direction that would have made the prediction better. This is repeated billions of times. The weights converge toward values that capture the patterns in the training data. After training, the weights are frozen and shipped as the model.

Weights are usually stored as 16-bit or 8-bit floating-point numbers. A 70B-parameter model at 16 bits takes about 140 GB on disk. Techniques like quantization compress those numbers to 4 bits or fewer to make models small enough to run on a laptop.

Examples

Open-weights release: Meta publishes Llama 4’s weights on Hugging Face. A developer downloads the file and runs it locally with no internet connection. The weights are exactly the same numbers Meta’s servers use.

Fine-tuning: A startup takes those weights and continues training them on their own customer-support transcripts. Most weights barely change; a small subset shift to encode the new behavior. The result is a customized model with the same architecture but different weights.

Quantized weights for laptops: The same 70B model is quantized from 16-bit to 4-bit weights. The file shrinks from ~140 GB to ~35 GB and now fits on a MacBook’s GPU memory, with a modest accuracy hit.

Sources

• Anthropic Transformer Circuits — mechanistic interpretability of weights
• Goodfellow, Bengio & Courville — Deep Learning, Chapter 6
• Hugging Face — Llama model card & weights

Last reviewed: May 2026

You May Also Like

• What is a Neural Network?
• What is a Large Language Model (LLM)?
• What is Fine-tuning?
• What is Quantization (in AI)?
• AI Glossary: 100+ Terms Every Beginner Needs to Know