What is an open-weight model?
A model whose trained numbers anyone can download. The mind is on your disk, which is the whole trick that makes local AI possible.
An open-weight model is one you can download, because its parameters are published rather than kept behind an API. You get the trained numbers and a license that lets you run them, on your own machine, for inspection, adaptation or anything the license allows. You do not get the training data or the training process. What the weights are meant to be good at is documented in the model card that ships alongside. Open weights are what every local tool runs on, and the term exists because they are not quite open source.
The weights are the model.
A language model is nothing but its parameters and a way to read them, so publishing the weights really does publish the model. Nothing else is held back at inference time: there is no server-side ingredient the model needs, no key, and no account. That completeness is why running a local LLM is a download rather than a service, and why an open-weight model keeps working on a machine that never touches the internet.
What you get, and what you do not.
Open weights buy you control. The model runs where your data is, its behavior does not change without a version you chose to install, and there is no meter between you and it. What you do not get is the recipe: the training data, the training code and the compute behind those weights usually stay private, which is why "open weight" is the honest term where "open source" would overpromise. The weights you do hold can still be adapted further; what fine-tuning is covers that road. The file format the community standardized makes the distinction concrete: you hold the finished artifact, not the kitchen it was baked in.
Why it carries local research work.
Grounded tools lean on open weights differently from chat tools. A chatbot needs the model's memory to be vast; a research notebook needs the model to read carefully, since the knowledge in the answers comes from your retrieved documents rather than from the model. Models in the 7B class handle that reading work well, which is why a personal machine is enough hardware and why the capability question shifts from what the model remembers to how well retrieval feeds it.