MiniMax-M2.7's MIT-Style License Is a Misleading Restriction That Bans Commercial Use and Fails Free Software Standards
Posted by pmttyji@reddit | LocalLLaMA | View on Reddit | 34 comments
Previous post(on this topic) by gnu.support
Really want this kind of posts for all custom licenses. Some(including me) couldn't custom licenses fully.
crantob@reddit
Poor minimax they just wanna make cool models and all the world of not-model-creators goes about fighting like mangy dingos over who gets to do what with the scraps of the kill.
Ok_Warning2146@reddit
I see. They make money from API, so they don't want people to take their business.
Objective-Picture-72@reddit
Minimax has updated their license. It's totally fine for 99% of applications.
wullyfooly@reddit
Where is the updated license. Not seeing it?
DistanceSolar1449@reddit
He’s lying, there’s only one license.
The history of the file is public lol
https://huggingface.co/MiniMaxAI/MiniMax-M2.7/commits/main/LICENSE (click on the “history” button)
EbbNorth7735@reddit
https://huggingface.co/MiniMaxAI/MiniMax-M2.7/blob/main/LICENSE
EbbNorth7735@reddit
https://huggingface.co/MiniMaxAI/MiniMax-M2.7/blob/main/LICENSE
Not clear if it still allows using the output of the model to create a product without licensing it first. It also lopks like it may still require branding any products made by minimax m2.7 as having been made by it. They need a clear average user breakdown of how it can be used.
Corporate_Drone31@reddit
Where is the link to the new license? It looks the same on HF.
Former_Basis3050@reddit
The worst part about these "custom MIT" licenses is exactly what you mentioned-they trick solo developers who don't have a legal team behind them. If there's a commercial restriction, just call it "Source-Available" or a "Research License" and be honest about it.
Slapping the word "MIT" on a restricted license just pollutes the open-source ecosystem and creates massive headaches down the line when someone accidentally builds a product on it. It's great that they are releasing the weights, but these licensing gymnastics are getting ridiculous. Thanks for posting this breakdown.
EbbNorth7735@reddit
Great article. One thing I don't understand is why not just add a clause that requires hosting restrictions to the model. If you use it for internal development and self host it it's free. If you pay for a service to use it it's free to you but make it so that you can't provide a service that utilizes it like a service provider or building it into your own products without a license. They would enable developers to utilize it and build with it and are more likely to build it into consumer facing products that give them revenue.
silenceimpaired@reddit
I think this is difficult to word legally. I think Black Forrest Labs tried to do this in their license and it was just confusing.
Still, as a not lawyer, I would phrase it something like this: This model can be used for commercial purposes provided all users are within 1 mile of the hardware it is being run on.
tomz17@reddit
How about simply "you can't charge users for access to it" ...
silenceimpaired@reddit
See even here… access to what? The weights? And by access do you mean like huggingface lets you download it?
I generally get what you’re driving at and hopefully they can come up with a plain language paragraph that basically says that… without leaving users wondering if they can profit off the output if they are the ones doing the hosting.
Black Forest Labs has an image model Flux that has a license that at least to me reads I can use the outputs of the model for commercial use provided I didn’t run the model myself. That’s what I fear.
tomz17@reddit
no, access as in "access to the model's output products"
silenceimpaired@reddit
Ah, but if I host a model on my own computer and take the output to make a book, I am giving access to the model’s output… does that mean I can’t sell the book?
Again I see your point and what you want to do because your intent is clear… but as soon as it is clear… I think that waters down their efforts to restrict the providers… so their lawyers push them towards less clear wording.
I hope I am wrong with all that said.
tomz17@reddit
That's not what the license says. That's also not what copyright law (at least in the USA) says. So yes, you *can* always sell the book (at least in the USA), even if you used a completely closed-weight model to produce it. The book just may not be protected by copyright if it is generated by AI instead of a human, but that's not a model-licensing issue.
Yes, you are grasping at straws.
EbbNorth7735@reddit
Obviously a lawyer can figure out the appropriate wording. No need to nitpick here
EbbNorth7735@reddit
Anything that generates revenue they require a cut.
L0TUSR00T@reddit
Not knowing who the author is, it feels a bit weird to go all out on the MiniMax's license as a self-proclaimed GNU supporter when pretty much every large model trains on all sorts of FOSS code and can replicate, say, GPL code without telling you doing so.
silenceimpaired@reddit
HOT TAKE - Just because a model trained on FOSS code doesn’t make it a violation of FOSS licensing ideals. That really falls on those who make a product from the model and fail to release under a FOSS model.
Besides all this, anyone who values AI automatically takes a collective position of sorts because… against the will of Copyright holders of software and works of writing their work is consumed with the end goal they and their work will no longer be needed once the model reaches the end goal.
For this reason I think it’s immoral and unethical for any AI model to be released under anything but a completely free license like Apache and MIT. At least with a license like this those who are taken advantage of can take advantage of the work that comes from their efforts.
L0TUSR00T@reddit
Well, that's kinda what I meant, and I think it shouldn't be a hot take (ideally). Though people here seem to be misunderstanding what I meant, so I'll rephrase it.
This guy seems to be supposing GNU, presumably its licenses including GPL and AGPL. These licenses are copyleft. MIT isn't. If the author brands themselves like GNU, MIT or not ultimately doesn't matter, and having a huge discussion on it feels strange.
By the spirit of the copyleft licenses (although not necessarily the laws I guess, sadly), the models using copyleft materials should be licensed as copyleft. This means the entire source code, not just weights, should be open. Really, it's philosophically so wrong to just take copyleft code and ignore it and even make the users to unknowingly do the same stuff.
suicidaleggroll@reddit
They’ve since clarified their stance. “Commercial”, as they were using it, means inference providers serving MiniMax to other customers. It does not mean companies using MiniMax to generate code for their own purposes. They’re going to be releasing an updated licensee with better wording soon.
One-Replacement-37@reddit
You should be grateful that a commercial company invests millions to train a model and then publishes open-source weights of a world leading model to the community.
Without them, this sub wouldn’t even exist and we’d all be paying hundreds to mega corps.
None of that is enforceable anyways - do you write “Implemented by minimax” everywhere? Do you even yourself publish any useful works for the community?
Stop complaining for a second.
silenceimpaired@reddit
Yeah, if someone wants to train on all copyrighted data without getting approval from the copyright holders… you should be grateful they think of you at all! They clearly don’t value the little guy.
Someone who did would at a minimum realize they have benefited from the collective works of many and release their models under Apache or MIT so everyone could benefit from them.
No, if someone slaps you without cause, you must ask for another and say thank you.
Uhlo@reddit
The problem here is not that the model was not published "open" enough. The problem is the framing "our license is MIT" when it clearly isn't. There are a lot of other open weight models with licenses that cleanly state "research license" or "personal license" or something. Calling it a "modified MIT" license and then being the opposite of MIT is deceiving.
But then again, your argument "if you don't publish a 230B parameter model yourself you have no right to complain about anything" is just bad faith. So I don't think you will (or want to) get my point.
ambient_temp_xeno@reddit
This sub exists because Llama 1 leaked.
LagOps91@reddit
is anything of that actually enforcable in the west? will they sue you? will the courts actually side with them?
Ok_Mammoth589@reddit
No it's not illegal to misrepresent things in the west. It's called corporate puffery
Uhlo@reddit
So if they would just call it "MIT License" I think there would be a strong case. It's not just what is written, but also what is expected - and if you call the license "MIT" then it should better be MIT, otherwise that is just deception. But calling it "modified MIT" is a bit trickier. But regardless, I am not a lawyer in any country on this planet or any other. So I'm just talking out of my ass. But still, it's just a very very bad move.
electroncarl123@reddit
API providers must be desperate to serve M2.7 - so much misinformation on "commercial" usage of the model.
ambient_temp_xeno@reddit
I got the feeling they weren't going to release it at all, so this is a glass half (maybe a quarter) full outcome.
One-Replacement-37@reddit
You should be grateful that a commercial company publishes open-source weights of a world leading model to the community. None of that is enforceable anyways. Without them, this sub wouldn’t even exist and we’d all be paying hundreds to mega corps. Stop complaining.
Uhlo@reddit
This is sooo misleading! At least call it the "MiniMax-Research-License" or something like the other companies are doing. Calling it MIT makes it sound open in a way it definitely is not. I would be interested what a lawyer would say to that license.
CATLLM@reddit
I'm afraid this misleading licensing thing is setting a precedence for future open-weights models and I don't like this timeline. 😭