Model weights are numerical parameters that define the behavior and knowledge of neural networks and language models. They are the learned values in each neuron and synapse across the layers of a deep learning model, stored as multidimensional arrays called tensors. Weights are adjusted during training to minimize prediction errors, and once training completes, they remain fixed until retraining occurs. For large language models, weights often number in the billions or trillions, representing the encoded patterns, facts, and reasoning capabilities the model has acquired.

Although model weights are often conceptualized as abstract digital data existing in “the cloud,” they are fundamentally physical arrangements of matter. Weights are stored as electrical charges in memory, magnetization on magnetic media, or transistor states in silicon chips. A large language model's weights can occupy 7–100+ gigabytes of storage, depending on model size and precision. These weights can be saved to hard drives, copied across networks, or physically transported to air-gapped (disconnected) systems via external storage devices¹⁾.

The physicality of weights has practical implications: they can be encrypted, version-controlled, backed up, and moved geographically like any other physical asset. Deploying a model requires transferring billions of individual numerical values to GPU or CPU infrastructure where computation occurs.

Weights determine how input data is transformed through the network's layers. During inference (prediction), input passes through matrix multiplications involving weights at each layer, producing outputs. The specific arrangement of weights encodes learned associations—for example, weights in a language model's attention layers capture relationships between words and concepts. Changing even a single weight slightly can alter model outputs, though larger weight perturbations cause more dramatic behavior shifts.

Fine-tuning modifies a subset of weights on new data, allowing models to specialize for particular domains or tasks without full retraining from scratch.

Modern approaches to model weight management include quantization—reducing precision from 32-bit floating-point to 8-bit or lower representations, decreasing storage and computational costs. Weights can also be distributed across multiple devices for parallel inference on very large models.

Different frameworks and hardware platforms represent weights in incompatible formats, so weight files often require conversion before deployment to different inference engines.

• Open-Weight Models
• What is an AI Model