What is an open-weight model?

In plain English

An AI model is, at heart, a huge list of numbers called weights. These are what the model learned during training, and they determine how it responds to any input. With most commercial AI, those numbers stay locked inside the company’s datacentre and you reach the model over the internet. With an open-weight model, the company publishes the numbers so anyone can download them and run the model on their own hardware or private cloud.

Open-weight is not the same as open source. The weights may be free to download while the training data, training code and full recipe stay private, and the licence may still restrict what you can do commercially.

Why it matters now

Two releases this week were open-weight: Mistral Large 4, which the company says it will publish by the end of the month, and Reflection’s Beam. Both are aimed at companies that want capability without dependency.

Running a model yourself means your data never leaves your systems, the model cannot be withdrawn or changed underneath you, and you set your own rules about what it will and will not do. Mistral makes that argument explicitly for cybersecurity work, where a closed model’s safety filters can refuse legitimate tasks such as proving a software flaw is real.

An example

A hospital wants an AI assistant to summarise patient notes. Sending those notes to an external service raises legal and privacy problems. With an open-weight model, the hospital’s IT team downloads the weights, runs the model on servers it controls, and the notes stay in the building.

What to watch out for

Downloading weights is the easy part. Running a large model needs serious hardware, and someone must handle updates, security and evaluation. Open weights also cannot be recalled: once released, they can be modified to remove safety restrictions, which is why labs typically run red-teaming before publishing.