AI Glossary

What is Open-Weight Model?

Insta's plain English

AI models you can download and run yourself, because the trained "weights" are public — think Llama, Mistral, or DeepSeek.

An open-weight model is an AI model whose trained parameters ("weights") are publicly released, so anyone can download, run, and adapt it — often for free.

The full picture

Most commercial AI is "closed": you access it only through the provider’s API and never see inside. Open-weight models are the opposite — the trained weights are published, so you can run the model on your own infrastructure, fine-tune it, and inspect it. Llama, Mistral, and DeepSeek are prominent examples.

Note the careful wording: "open-weight" isn’t always fully "open-source" (training data and code may stay private), and licences vary. Still, open-weight models are reshaping the market — offering cost control, privacy, and independence from any single vendor, and letting capable AI run even where data can’t leave the building.

📌 Real business example

A healthcare firm that can’t send patient data to an external API self-hosts an open-weight model inside its own network, keeping sensitive information in-house while still using modern AI.

How different roles use this

Technical lead

Self-hosts or fine-tunes open-weight models for privacy, cost control, and customisation.

Business owner

Weighs open-weight options to reduce per-call API costs and avoid vendor lock-in.

Compliance lead

Uses self-hosted open-weight models to keep regulated data on-premise.

Common questions

Q: Is open-weight the same as open-source?

Not exactly. Open-weight means the trained model is downloadable. Fully open-source would also include the training code and data. Always check the specific licence.

Q: Why choose an open-weight model?

Control and cost: you can run it on your own hardware, keep data private, fine-tune it freely, and avoid depending on one vendor’s API and pricing.