1. Home
  2. Glossary
  3. Model weights
AI glossary · Models & training

Model weights

Model weights: Model weights are the numbers inside a trained neural network that determine how it responds to input. They are what training produces and what a company ships or protects; releasing them publicly is what makes a model 'open weights.'

A neural network is a structure of connected units, and each connection has a numeric strength called a weight. Training adjusts those numbers, millions or billions of them, until the network's outputs match the desired ones. Once training is done, the weights are the model. Copy the file of weights and you have a working copy of the model.

That is why weights are treated as crown jewels. Companies such as OpenAI, Anthropic, and Google keep the weights of their flagship models private and offer access through apps and APIs. Others, including Meta with its Llama models and Mistral, publish weights so anyone can download and run them on their own hardware, subject to a license.

For a business, the distinction matters for control and privacy. Running open-weights models on your own servers keeps data in-house and lets a technical team fine-tune freely, at the cost of managing infrastructure. Using a proprietary model through an API is simpler and often more capable, but your data passes through the vendor under its terms.

Example at work

A healthcare software company handling protected health information decides to run an open-weights model inside its own cloud environment so patient data never leaves its control. It accepts somewhat lower capability than the top proprietary models in exchange for a simpler compliance story with its hospital customers.

Why it matters

When a vendor talks about 'open' models, self-hosting, or fine-tuning, they are talking about weights. Understanding the term lets you follow those conversations and weigh the tradeoff between convenience, capability, cost, and control that every AI deployment involves.

Related terms