Open weights, closed weights, and what “open” is hiding

Three things can be open

A trained model, you learned in the first module, is a file of numbers. Making it is a process with three parts: the weights that result, the code that trained them, and the data they were trained on. Any of the three can be published or withheld. The phrase “open source” was borrowed from software, where there is one thing to publish, and it fits badly here, because most “open” models publish one part of three.

In practice there are three tiers.

Closed. You send text to a server and get text back. You cannot download the model, inspect it, or run it anywhere but where the lab runs it. The leading models from the largest American labs are closed.

Open weights. The file of numbers is published and anyone can download and run it. The data is described in general terms and not released; the training code usually is not either. This is what almost every “open” model means.

Fully open. Weights, code and the data itself. A handful of research models, notably from a non-profit institute in Seattle since 2024, meet this bar. Nobody at the frontier does.

How the weights got out

The first large open-weights model was not meant to be one. In February 2023 Meta released the weights of its research model to approved researchers under a licence. Within a week the file was on a public torrent. Meta’s response, that July, was to release the successor openly on purpose, with a licence attached.

The licence is where “open” gets complicated. It lets you use and modify the model, but a company with more than seven hundred million monthly users needs separate permission, and for a time you could not use its output to improve a competing model. Those are restrictions, and open-source software licences do not have them. A French lab that September released a model under an unrestricted software licence, and several Chinese labs have since done the same; the model that caused the stock-market reaction in early 2025 was released under the most permissive licence in common use.

In October 2024 the body that maintains the definition of open-source software published a definition for AI. It requires the weights, the code, and enough information about the data to rebuild a substantially equivalent model. Meta’s models do not meet it. Meta continues to call them open source. The disagreement is not resolved and you should expect the word to keep being used both ways.

Why a lab gives it away

It is not charity, and understanding the reasons tells you how stable the supply is. A company whose business is not selling models benefits when models are cheap and abundant, because that makes its actual products more valuable. Releasing weights attracts researchers and builds an ecosystem of tools around your design rather than a rival’s. Governments in China, France and the Gulf treat open models as national capability. And a model that many people run and study is harder to regulate out of existence.

Every one of those reasons can change. A lab that releases openly today can stop tomorrow, and the models it already released cannot be recalled.

What it gives you

Four things, and each connects to a limit met earlier in this course.

You can run it on your own hardware, which a later module does, and which means your words never leave your machine. You can adapt it to your own data. It cannot change under you: the lesson on the model that changed under you does not apply to a file you hold. And researchers can look inside it, which is how much of what is known about how these models work was found.

For India specifically, open weights are the base under most of the country’s own language work: the translation systems for twenty-two scheduled languages from a group at IIT Madras, and several Indian labs’ Hindi and Indic models, started from open weights and added Indian text.

What it does not give you

You cannot inspect the data, so you cannot know what the model read, what it was filtered for, or whose text is in it. You cannot verify a lab’s claims about it. The safety training can be removed (a later module explains why in a few hundred lines of code), so an open model is also an open model for anyone. And open models trail the closed frontier; the gap was estimated at about a year in 2024 and narrowed to months in 2025, but it has not closed.

The right question about any model described as open is the one this lesson started with: open which part, under what licence, and what has the lab chosen not to say.