What is Open Weights? — one analogy

1The analogy

Buy a scent you like at a perfume counter and you get the whole bottle. Take it home, spray as much as you want, decant it into a smaller bottle, or mix in a few drops of something else to make it your own. The bottle is now completely yours.

But the shop does not hand over the formula. It will not tell you which ingredients came from where, in what order, or in what amounts. A few ingredient names printed on the back are not enough to recreate the same scent.

So if you ever wonder why it smells the way it does, there is no way to find out. All you can do is spray it and judge for yourself.

Open weights is that bottle of perfume.

2In detail

The numbers are the model

A trained model, in the end, is one enormous table of numbers — which value sits where, and how much of it. That table is the entirety of what the model knows. Bundling that table into a file and putting it out there is open weights.

Having the file does not mean the inside is transparent. None of the individual numbers carries a label. No person can read off which value handles which piece of knowledge. "Open" here means less "you can look inside" and more "you can take it and run it."

Some things are visible straight from the file, though. What structure it is built on, roughly how big it is, and roughly what hardware it takes to run — those show up as soon as you open it.

What you can do once you have it

The biggest difference is where it runs. Borrowing someone else's window means your text has to pass through their computer; having the file in hand means you can run it on your own turf. That difference matters a lot when you are handling material that should never leave the building.

Running without an internet connection is another advantage. Even if the service that made it shuts down, or prices go up, or something quietly changes, the file you already downloaded stays exactly as it was. Today's answer and yesterday's answer stay closer together.

You can also tweak it. Like mixing in a few extra drops of a different scent, you can train it a little further on material from your own field to adjust its tone and fit. Shrinking it down to fit a small device is another thing you can only do once the file is in hand.

Where it splits from an open-source model

The names sound alike, so people mix them up, but what gets opened is different. An open-source model opens the build procedure and the material it used, so others can follow along and rebuild it. Open weights opens only the finished file.

That difference turns into a question of verification. If you do not know what material a model learned from, it is hard to say why a result leans one way. There is no way to check whether someone else's writing went into training, or whether personal information slipped in. Using it and judging for yourself is the only path there is.

The terms attached to the file matter too. Even with "open" in the name, some come with conditions on use or scale. Checking those conditions is the first step before putting one to work.

Once it spreads, there is no pulling it back

The moment a file goes out into the world, it gets copied endlessly. Even if a problem turns up later, what has already gone out cannot be recalled, because it keeps running wherever it was downloaded.

That trait carries a good side and a worrying one together. It is a good thing that a model outlives the team that made it, even if that team disappears one day. On the other hand, a stripped-down copy with its safety measures removed is just as hard to stop, because anyone holding the file can edit what is inside.

So the teams that release these files do not just release them blind. Many pair the release with a document describing what it is good at, where it is weak, and what risks it carries, and open things up in stages rather than all at once.

3More precisely

In "open weights," the weights are the values fixed by training. A model only runs once those values are paired with the structure that defines what order to use them in. That is why a release usually ships the value file and a description of the structure together — values alone, without the structure, do nothing.

The perfume comparison breaks down somewhere too. Sharing perfume shrinks your own share; copying a file does not shrink the original at all no matter how many times it happens, which is why a file spreads far faster than perfume ever could. Perfume also changes its scent the moment you mix in a few drops, while editing a model takes serious hardware and time. There is still a real threshold between having something in hand and actually being able to change it freely.

How much gets opened is not fixed either. Some teams release only the file. Others pair the file with a document on strengths and weaknesses. A few even disclose what kind of material went into training. The same word covers very different amounts of actual transparency.

4Try it yourself

5Common misconceptions

It's easy to think open weights and open source are the same thing, but actually only the finished file gets opened here, and the build method stays closed.
It's easy to think an open file means you can see inside it, but actually it is an unlabeled block of numbers that is hard for a person to read.
It's easy to think you can do anything once you've downloaded it, but actually some files carry conditions on use or scale.

7One-line summary

In shortOpen weights releases only the trained set of numbers for anyone to download, so you can run it and adjust it, but you still cannot know exactly how it was built.

Did this analogy help?

Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02