layman/terms
AI words

Open-weights model

oh · pun wayts mod · ul

In shortAn open-weights model is an AI model whose finished settings anyone can download, so they can run it on their own computer instead of only using it through a company's app.

Some bakeries only sell you the cake. Others pin the recipe to the wall, measurements and all.

A bakery can sell you a cake. Or it can pin the recipe to the wall with every measurement: how much flour, how many eggs, how long in the oven. With the recipe you can bake the same cake in your own kitchen.

An open-weights model is an AI model whose recipe is pinned to the wall. Anyone can download its finished settings, called weights, and run it on their own computer.

What the weights are

When a model is trained it ends up with billions of numbers, its weights. They hold everything it learned. Some companies keep their weights secret and only let you use the model through their app. Others publish them for anyone to download.

Why people want them

The catch

The biggest, smartest models are often closed. Open-weights ones can be a step behind and the large ones need a powerful computer to run. You also look after it yourself: no company is fixing it for you.

Where the comparison breaks down

A recipe also tells you how to make it. Open weights usually don’t come with the training data or the full method, only the finished result. That’s why some people say it isn’t truly open source: you get the cake’s exact measurements, not the farm the flour came from.

How the words fit together

An open-weights model is an LLM you can download. Running it is inference, the cheap part; making it was training, the expensive part someone else already did. You can fine-tune it to fit your own work.

Watch out

Open weights isn't always the same as open source. Check the licence before you use one at work: some only allow personal use or set limits for big companies.

When you'll hear it

“We can run it ourselves.” The model is open-weights, so it can sit on our own computers and our data never leaves.

“It's a 7B model.” It has about 7 billion settings, small enough for a good laptop.

“We fine-tuned Llama.” We took an open-weights model and taught it a bit more about our own work.

Say it in a meeting

“Does our data need to stay on our own machines, or is a hosted AI app fine?”

Try it yourselfReady to use it? The exact commands are on my other site, devcmds: Run an AI model on your computer ↗