AI Glossary

What is Model Weights?

Insta's plain English

The trained numbers that *are* the AI model. Hold the weights and you can run or copy it — so protecting them is a real security issue.

Model weights are the numerical values a model learns during training — effectively the trained "brain" of an AI. Whoever holds the weights can run, copy, or fine-tune the model, which is why their security matters.

Defined in the Great American AI Act (discussion draft). Read it in context →

The full picture

When an AI model is trained, the result is a large set of numerical parameters — the weights. They encode everything the model learned, and a copy of the weights is a copy of the model’s capabilities. "Open-weight" models publish them; closed models guard them closely.

Because the weights are the crown jewels, their security is now a policy and business concern: frontier-AI rules ask large developers to address "model weight cybersecurity" — protecting weights from theft or leak. For businesses, this matters when self-hosting open-weight models (you’re responsible for securing them) and when assessing a vendor’s protection of the models you depend on.

📌 Real business example

A company self-hosting an open-weight model treats the weights file like sensitive IP — access-controlled and encrypted — rather than a normal asset, since anyone who copies it gets the full model.

How different roles use this

Technical lead

Secures self-hosted model weights (access control, encryption) as high-value assets.

Security lead

Assesses "model weight cybersecurity" — theft/leak risk — for both vendors and in-house models.

Business owner

Understands that holding the weights means holding the model, which informs build-vs-buy and IP decisions.

Common questions

Q: Why do model weights need protecting?

The weights are the model. A stolen copy hands over the full capability — and, for powerful models, that’s a security risk, which is why regulation calls out model-weight cybersecurity.

Q: Are weights the same as training data?

No. Training data is what the model learned from; weights are the learned result. You can have the weights without the original data.