Core Concepts
Model Weights in plain English.
Also known as: weights,neural network weights
The one-sentence version
The millions or billions of learned numbers inside a trained AI model that encode everything it knows.
Model weights are the millions to trillions of numerical parameters inside a trained neural network that encode everything the model has learned — adjusted repeatedly during training until the model gets good at predicting text, images, or whatever it was trained on. When people talk about a model having "70 billion parameters," they mean 70 billion of these numbers. Weights are the actual product of training: enormously expensive to produce (frontier training runs cost hundreds of millions of dollars) but, once produced, cheap to copy and run. Whether a company releases its weights publicly is the line between "open-weight" models like Llama 4 and DeepSeek V3, which anyone can download and run, and "closed-weight" models like GPT-5.6 and Claude Opus 4.8, which are only accessible through an API the company controls. Leaked or stolen weights are treated as a serious security concern, since they hand over the trained model itself, not just access to it.