Weights

In a neural network, weights decide how much influence the input will have on the output.

How it works

Weights are the learnable numerical parameters of a neural network stored as floating-point numbers in matrices and tensors. Each weight represents the strength of a connection between two neurons. During training, backpropagation computes how each weight contributes to the prediction error, and the optimiser adjusts each weight by a small step in the direction that reduces that error. After training, the weights are frozen; they encode everything the model has learned from the training data.

Why it matters

When people talk about a '70B parameter model,' they are referring to 70 billion weights. These weights are the model — they are what gets downloaded when you pull an open-weight model, what occupies GPU VRAM during inference, and what gets updated during fine-tuning. The entire business model of open-weight AI (LLaMA, Mistral, DeepSeek) is built around sharing these trained weight files. Understanding weights clarifies why model size, memory requirements, and fine-tuning costs are all directly related.