OpenAI publishes open weights now

For years the position was simple: releasing model weights was dangerous, and the company that did it most was the one most often criticised for it.

Then OpenAI released the GPT-OSS family — and did it under Apache-2.0, the licence that asks for nothing.

What shipped

Two models, both genuinely open:

• GPT-OSS 20B — about 20 billion parameters, of which roughly 3.6 billion are active per token
• GPT-OSS 120B — around 120 billion, with about 5.1 billion active

Apache-2.0 means you can use them commercially, modify them, redistribute them, and build a product on them without asking anyone or paying anything. It is the same licence as most of the web's infrastructure.

Why the number of active parameters matters more than the total

Both are mixture-of-experts models. Only a fraction of the network runs for each token, which is why the 20B model fits comfortably on consumer hardware while behaving like something much larger.

It is the single most important development in local AI in years: capability without a proportional memory bill.

Why they did it

Three plausible reasons, and probably all three.

The floor moved. Other labs had spent two years giving away capable models. Once a 30B-class open model exists, withholding your own buys nothing but bad press.

Developers are the prize. Someone who builds on your weights tends to build with your tools, your format and your habits. The model is a loss leader for an ecosystem.

It is no longer the frontier. These are strong models. They are not the company's best models, and the gap is the point — you can give away last year's capability without giving away this year's.

What it changes for you

Quite a lot, if you were on the fence about running a model yourself.

A permissively licensed model at this size means you can build something and sell it, without the licence questions that hang over most of the field. Several of the most popular open-weight models carry non-commercial terms or limits on scale. These do not.

It also means the honest answer to "can I run a capable model on my own machine?" is now yes, and it stopped being a compromise.

The caveat worth keeping

Open weights are not open source. The licence here happens to be genuinely permissive, which is unusual and worth noting — see why most open AI is not open source.

But the weights are downloadable and the licence has no conditions. For once, that is not a hedge.