What is a Model Weights?

A weight is a single number that scales how much influence one neuron's output has on the next layer; a large language model has billions of these numbers arranged across its layers. Training is the process of adjusting weights via gradient descent so that the model's outputs increasingly match the desired training signal; inference is simply running a fixed, unchanging set of weights forward over new input. Weights are typically distributed as checkpoint files (safetensors, GGUF, PyTorch .bin) that can be loaded into inference software to reproduce a model's exact behavior. Whether weights are published openly (open-weights models like Llama or Mistral) or kept private behind an API (closed models like GPT or Claude) is a major dividing line in the AI ecosystem, affecting who can self-host, fine-tune, or audit a given model.