Hacker News Reader

Comment by kennywinker | Hacker News Reader

I understand the sticklers who are annoyed at people using open weight and open source interchangeably, but it misses some nuance. A common refrain I see is "it's like saying a shareware app is open source!" - that's just wrong.

Open source means you can modify, rebuild, and re-distribute the code and the resulting app. Open weights you can do two of those things - you can modify and re-distributed. They can't be re-built, since you're missing the original training data - but they are a form of open that exists somewhere between open source and closed source. Derivative works are allowed, but the source material isn't available.

I don't see people complaining about creative commons licensed images or audio being "not real open source", and open weight models seem about the same as that to me.

View on HN

I disagree. Open source is when the source is open. Free software is when you are free to make modifications and redistribute.

Edit: as bayindirh noted, according to OSI's definition, I am completely incorrect about redistribution of modifications.

The definition of Open Source [0] is freer than what you portray, but Open Weights is not open by this definition. As I said in the thread, it's more like "weights available, with some freedoms attached".

Current state is detrimental to both Free/Open Software and AI ecosystems at the same time.

[0]: https://opensource.org/osd

I stand corrected.

My point was never that open weights = open source, so it’s not surprising that they fail the OSI definition of open source.

My point is that open weights exists somewhere on the spectrum between closed source and open source, and imo much closer to open source than closed.

As evidence, see the thousands of open-weight model derived models on hugging face.

you're still correct.

"open source" is ambiguous (in cases this is deliberate)

OSI has its own definition of open source, but there's no reason someone can't say their code is "open source" because you can see the source code.

This is akin to saying 10tb hard disk... very few people quantify this with tera or tebi or Ti or whatever unless there is some legal fineprint required. and there is no hard legal definition for calling something "open source".

In any case, stallman has a take on the difference in philosophy of the OSI version:

https://www.gnu.org/philosophy/open-source-misses-the-point....

If you couldnt tell I am very much on the Stallman side of things, but I almost wish they had been "libre" from the start so you don't have to constantly make the free beer/speech distinction when talking about it.

> I don't see people complaining about creative commons licensed images or audio being "not real open source", and open weight models seem about the same as that to me.

Creative Commons licenses are a spectrum from CC BY to CC BY-NC-ND. All six licenses offer different freedoms and requirements about the work licensed.

As I said in another comment, Open meant that you can buy an interface to the thing you are interested in. Nothing like Open Source or Free Software of today, and Open Weights lean to the former than the latter, this is what rubs the people the wrong way (incl. me).

A better term would be weights-available. Do not create false assumptions about so-called openness of a model, and is akin to sharing an object file which you can compile into something but its source is not available to start from scratch.

Nuances matter, that's true, but the packaging of the facts also matter. A good example is, again, XKCD's comic titled Free [0]. Open weights are open as the open of 2000s commercial software, maybe with a free SDK, but the labeling is deceiving on purpose.

[0]: https://xkcd.com/641/

> Open weights are open as the open of 2000s commercial software, maybe with a free SDK

This just isn’t true. 2000s commercial software you buy under a restrictive license, and then extend using an sdk. You can’t modify or distribute the code only your plugin or extension. Open weight models you freely download and modify and extend using open source tools, and create derivative works. The licenses tend to be very permissive (e.g. Apache 2.0 https://huggingface.co/Qwen/Qwen3.8-27B/blob/main/LICENSE)

Perhaps you are not aware of the thousands of derivative models that have been created based on open weight models? Ones tuned for this or that task, expanded, hacked apart.

> akin to sharing an object file which you can compile into something

I see what you are getting at, but modifying object files isn’t practical or simple or commonly done. A model that’s been post trained isn’t the old model with a lump of new code attached to it. The weights have been irreversibly altered and it’s a new thing now. A derivative work.

My contention is that model weights are a form of source code. They the software AND a modifiable thing, closer to minified js than a binary blob. Less flexible than the full training data would be, in some ways, but more useful in others (unless you have a large data center at your disposal). Would you consider something closed source if it had a bunch of generated code, but didn’t include the code used to generate the generated code? I would.

Reply on news.ycombinator.com