FSF: Meta's License for Its Llama 3.1 AI Model 'is Not a Free Software License' (fsf.org) 35
July saw the news that Meta had launched a powerful open-source AI model, Llama 3.1.
But the Free Software Foundation evaluated Llama 3.1's license agreement, and announced this week that "this is not a free software license and you should not use it, nor any software released under it." Not only does it deny users their freedom, but it also purports to hand over powers to the licensors that should only be exercised through lawmaking by democratically-elected governments.
Moreover, it has been applied by Meta to a machine-learning (ML) application, even though the license completely fails to address software freedom challenges inherent in such applications....
We decided to review the Llama license because it is being applied to an ML application and model, while at the same time being presented by Meta as if it grants users a degree of software freedom. This is certainly not the case, and we want the free software community to have clarity on this.
In other news, the FSF also announced the winner of the logo contest for their big upcoming 40th anniversary celebration.
But the Free Software Foundation evaluated Llama 3.1's license agreement, and announced this week that "this is not a free software license and you should not use it, nor any software released under it." Not only does it deny users their freedom, but it also purports to hand over powers to the licensors that should only be exercised through lawmaking by democratically-elected governments.
Moreover, it has been applied by Meta to a machine-learning (ML) application, even though the license completely fails to address software freedom challenges inherent in such applications....
We decided to review the Llama license because it is being applied to an ML application and model, while at the same time being presented by Meta as if it grants users a degree of software freedom. This is certainly not the case, and we want the free software community to have clarity on this.
In other news, the FSF also announced the winner of the logo contest for their big upcoming 40th anniversary celebration.
They have a presentation at Fosdem on 2 Feb (Score:5, Informative)
FSF's Zoe Kooyman and Krzysztof Siewicz will give a presentation on Sunday 2nd of Feb:
"FSF's criteria for free machine learning applications"
https://fosdem.org/2025/schedu... [fosdem.org]
It'll be streamed. Well worth tuning in for. A recording should be online soon after.
Free vs. Open Source (Score:3, Insightful)
Re:Free vs. Open Source (Score:5, Informative)
FSF does not have a definition of Open Source. They do not comment on if Llama is Open Source. They are just pointing out that it is not Free Software.
Re: (Score:3)
FSF does not have a definition of Open Source. They do not comment on if Llama is Open Source. They are just pointing out that it is not Free Software.
I agree, however TFA refers to Llama as open source, not free. FSF certainly can say it doesn't meet their definition of free. Meta's license certainly is restrictive, but they do not claim it is free software as defined by the FEF, either.
Re: Free vs. Open Source (Score:3)
the OSI defines open-source, and it's unlikely that license would comply either.
Re: (Score:2)
You're getting confused.
Llama does not claim to be open source. It only claims to be open weights.
Re: (Score:2)
You're getting confused.
Llama does not claim to be open source. It only claims to be open weights.
The home page header at Meta's llama.com has:
The open-source AI models you can fine-tune, distill and deploy anywhere. Choose from our collection of models: Llama 3.1, Llama 3.2, Llama 3.3.
It's pretty clear they claim their models are open source.
It's published by Meta (Score:5, Interesting)
Re:It's published by Meta (Score:5, Insightful)
Well let me save you the trouble then!
The objection here is on the 25 prohibited usages, which FSF believes should be handled by the local laws of your jurisdiction.
e.g. 'the harassment, abuse, threatening, or bullying of individuals or groups of individuals'. Noble sentiments but unenforceable, are you gonna get a threatening court order from the morality police at Facebook legal because a user of your software posted something mean about someone else online?
(In case it weren't clear, I'm not condoning harassment or bullying; "Be excellent to each other".)
Re: (Score:2)
^^^
I wish I had mod points because that was both informative AND insightful.
Re:It's published by Meta (Score:5, Interesting)
Noble sentiments but unenforceable
That's not the point.
If someone uses Meta's software to bully or harass, the victim will sue Meta (deep pockets) rather than the bully (broke).
The purpose of the restrictions is to give Meta legal cover to say the bully wasn't authorized to use the software.
Re:It's published by Meta (Score:4, Interesting)
You're quite correct.
In the old days a judge would have dismissed such claims with prejudice but in the 1960's the lawyers took over the courts and now almost all judges were lawyers and they are there to make other lawyers rich.
We pay for this corrupt judicial tyranny in the cost of every good and service.
Both FSF and Meta are correct but FSF assumes a fair legal system which is incorrect.
Re: (Score:2)
Wouldn't that be covered by clauses requiring the user to follow laws in the use of the software and indemnifying Meta for lawsuits? Clauses like that are already in the Lama license.
Re: It's published by Meta (Score:2)
Re: (Score:3)
It's also overly broad. You could have the software point out that Elon did a gesture commonly associated with 1930-1940s Europe and that would be harassing him.
Re: (Score:2)
However, these accusations were based on appearances alone and ignored the context of Musk's words.
During his brief speech — available in full on YouTube — Musk said to the crowd gathered in honor of Trump's victory, "And I just want to say thank you for making it happen. Thank you." He then briskly extended his right hand from his heart to the crowd in front of him. He repeated the gesture for the people behind him. Then, in a moment omitted from the users' videos, he said, placing his hand ove
Re: (Score:2)
Thank you for showing how it would be argued to be harassment.
Re: (Score:2)
Musk did a Nazi salute, twice.
The Nazi salute isn't "associated with Europe in the 1930-1940s", it is associated with the Nazi regime and its fans, back then and today.
Re: (Score:2)
I am well aware; I was pointing out that even a watered-down reference would be called harassment - thus the restriction is overly broad.
Re: (Score:2)
Ok, my misunderstanding.
Re: It's published by Meta (Score:2)
Re: (Score:2)
I'm pretty sure neither you, nor the space Nazi were around before that year in order to get so attached to it. Especially the space Nazi, which didn't even get to the US until the early 1990s.
I'm also fairly sure both of you know the reason it fell into disuse in 1942 .
Re: (Score:2)
The point for the FSF is: Freedom 0 is "I am allowed to use the product for any purpose I want to use" and that includes all things on the do-no-evil list. ..." is, that laws and not licensed should handle that.
The reasoning when someone says "But it isn't ethical to do
This does not only mean that people releasing software (AI models, etc.) shouldn't be required to decide what acceptable use is, but also that laws can be changed when new issues arise while most licenses cannot be retroactively changed.
Re: (Score:2)
There's also the "if you're big enough to complete with us, you can't use it at all"
If, on the Llama 3.1 version release date, the monthly active users of the products or services made available by or for Licensee, or Licensee’s affiliates, is greater than 700 million monthly active users in the preceding calendar month, you must request a license from Meta, which Meta may grant to you in its sole discretion, and you are not authorized to exercise any of the rights under this Agreement unless or until Meta otherwise expressly grants you such rights.
No doubt they picked that number to exclude all current competition.
No OpenAI, no Microsoft, no Google, no Apple, no Bytedance, no Canva, no Adobe...
It is not and everybody knows (Score:2)
Look at the related reddits and you see how everyone accuses the large model makers if they do not use Apache 2 or MIT (or similar). Some outright refuse to even test models with licenses like Llama 3.1 or Gemma have.
Currently there are a few high quality open models like Mistral-Nemo (Apache 2) and Phi-4 (MIT), but Mistral-Small (misleading name as it is larger than Nemo) has the Mistral research license that is non-commercial. The interesting thing is, that many of the Chinese Models have weak copyleft li
Re:It is not and everybody knows (Score:5, Interesting)
Re: It is not and everybody knows (Score:3)
Only when using these models from service providers. When the same models are run locally, they work well.
Re: (Score:2)
Yes, they are censored. The usual ways around censorship work in part, but it also seems that the trained data has been filtered, which means that some information is simply not there, even if there are no refusals from the model.
Wait, what? (Score:3)
Meta is misrepresenting something? Say it ain't so!
Can't "help" people (Score:2)
We want everyone to use Llama 3.1 safely and responsibly. You agree you will not use, or allow others to use, Llama 3.1 to:
Violate the law or others’ rights, including to:
Self-harm or harm to others, including suicide, cutting, and eating disorders
People have a right to self-harm and you can't use Llama 3.1 to interfere with that.
Hidden Traps in Meta’s Llama License (Score:1)
There is an article explaining that not only is Llama not open source, but it is also a license with a very high risk.
Meta can update the AUP at any time, and it is possible to trap all users. In addition, this trap is also effective for users of services that use Llama.
https://shujisado.org/2025/01/... [shujisado.org]