This is why I've always considered the weights-vs-source debate to be an enormous red herring that skips the far more important question: are the weights actually "Open" in the first place?
If Llama released everything that the most zealous opponents of weights=source demand they release under the same license that they're currently offering the weights under, we'd still be left with something that falls cleanly into the category of Source Available. It's a generous Source Available, but removes many of the freedoms that are part of both the Open Source and Free Software Definitions.
Fighting over weights vs source implicitly cedes the far more important ground in the battle over the soul of FOSS, and that will have ripple effects across the industry in ways that ceding weights=source never would.
I don't think most people in the weights-vs-source debate misunderstands this, it's just the the current "open-source" models for the most part do not even meet the bar of source-available, so talking about if the license is actually Open is not the current discussion.
See, but my point is that this is putting the cart before the horse. The "Open" in "Open Source" is what matters most by far, the same way that the "Free" in "Free Software" is the key word that qualifies the kind of software we're taking about.
Once we've resolved the problem of using the word "Open" incorrectly I'm happy to have a conversation about what should be the preferred form for modification (i.e. source) of an LLM. But that is the less important and far more esoteric discussion (and one about which reasonable people can and do disagree), to the point where it's merely a distraction from the incredibly meaningful and important problem of calling something "Open Source" while attaching an Acceptable Use policy to it.
> The "Open" in "Open Source" is what matters most by far, the same way that the "Free" in "Free Software" is the key word that qualifies the kind of software we're taking about.
I don't think this is true. If someone said "look, my software is open source" and by "source" they meant the binary they shipped, the specific definition of "open" they chose to use would not matter much for the sort of things I'd like to do with an open source project. Both are important.
I agree that both matter, but one is more important than the other.
If they released the binary as "Open Source" but had a long list of things I wasn't allowed to do with it, the fact that they didn't release the source code would be of secondary concern to the fact that they're calling it "Open" while it actually has a trail of legal landmines waiting to bite anyone who tries to use it as free software.
And that's with a clear cut case like a binary-only release. With an LLM there's a lot of room for debate about what counts as the preferred form for making modifications to the work (or heck, what even counts as the work). That question is wide open for debate, and it's not worth having that debate when there's a far more egregious problem with their usage.
The catch is that the benefits of open vs non-open don't translate neatly from software to models. If software is binary-only, is it exceedingly difficult to change it in any kind of substantial way (you can change the machine code directly, of course, but the very nature of the format makes this very limited). OTOH with a large language model with open weights but without open training data - the closest equivalent to open source for software - you can still change its behavior very substantially with finetuning or remixing layers (from different models even!).
> OTOH with a large language model with open weights but without open training data - the closest equivalent to open source for software - you can still change its behavior very substantially with finetuning or remixing layers (from different models even!).
The closest thing to open source would be to have open training data. The weights are the binary, the training date is the source and the process of getting the weights is the compilation process.
Fintuning or whatever is just modding the binaries. Remixing different layers is creating a workflow pipeline by combining different functions of a binary software package together with components from other binary software packages.
> and by "source" they meant the binary they shipped
Common misconception. Weights are not binary. Weights are hardcoded values that you load into an (open-source or closed-source) engine and you run that engine. The source code for LLMs is both in the architecture (i.e. what to do with those hardcoded values) and the inference engines.
As opposed to binaries, you can modify weights. You can adjust them, tune them for downstream tasks and so on. And more importantly, in theory you the downloader and "company x" the releaser of the model use the same methods and technologies to modify the weights. (in contrast to a binary release where you can only modify only the machine language while the creator can modify the source-code and recompile).
LLamas aren't open source because the license under which they're released isn't open source. There are plenty of models that are open source tho: mistrals (apache 2.0), qwens (apache2.0), deepseeks (mit), glms (mit) and so on.
what you describe reminds me pretty much of a binary blob that is loaded into a machine or software.
additionally modifying data in binary form was a longtime practice last time I looked, but I might not remember correctly.
In today's world, if Meta did release full source they used to create Llama, there are only about a dozen institutions that have the capacity to actually do anything with that, and no one has that kind of spare capacity just lying around. So the question of having the source for now in this case is less about being able to do something with it, and more about behind able to examine what's going into it. Aside from making it so it won't tell me how to make cocaine or bombs, what other directives has it been programmed with on top of the intial training run. That's what's important here, so I disagree that is a red herring. Both aspects are important here, but the most important one is to not let Mark Zuckerberg co-opt the term Open Source when it's only model available, and definitely not even actually Open at that.