TildeAI

/

TildeOpen-30b

like 133

Follow

Tilde 90

I am not sure if I understand your concerns
It has nothing to do with Llama or any other model. TildeOpen is trained from scratch.
CC-BY-4.0 is widely accepted for AI model weights and allows full commercial use, modification, and redistribution, which aligns with our goal of supporting underrepresented European languages.
Are you sure you know what CC-BY-4.0 is?

Yes, this response makes sense. Companies like Nvidia, Meta, Huyuan, etc are indeed using this, and definitely much, much more egregious licenses! In fact, most "community/open licenses" would really definitely leave you in a place where you sentiment is kind of the only logical one from a lot of angles.

However, copyleft, libre not gratis, the OSI's definition of OSS, and free software definitely have... quite a bit of learning to do in this arena if you are legitimately interested! Especially when dealing with sensitive data and intelligence, while it can be hard to trace the logical implications that would make such a license unfit, the problem is the broader ripple effect of using licenses that are not TRULY permissive in spirit AND function.

If that is the company this project would like to keep... by all means. Not a good look, in my opinion, but hey, Nvidia rules the market value so clearly there are many people with different opinions than I, but there is a reason the others use apache-2.0 and mit, and sometimes GNU/GPL. There is a reason more copyleft licenses than even those exist. This is the reason. Why would everyone be using those if this is fine? Think.

K2's modified MIT was minimal. However, it's not really that long after, there's another open source 1t parameter model that is truly open, and better agentic coders exist. Despite the novel and incredibly expensive training run... there is literally zero incentive to build anything on the platform, because there are truly option options that surpass it that were made from scratch despite K2 existing. Even still we saw finetunes, but imagine what the platform might have been. We'll never know. Why? Because the license was less permissive. It doesn't matter if it actually restricts or does anything. Workplaces and labs REALLY DO require the parts of the stack you're working with to be fully permissive sometimes, one of the accepted licenses from the list or go kick dirt. It means this model is completely invisible to many legitimate use cases. I have personally worked jobs where a plan was denied for use of CC-by-4.0 license in a technology inside of the stack, so yes... I know what it is. That cost me real money in planning time and work I had to do over.

Even if you fully disagree with me and think I'm an idiot, I guarantee I'm not the only idiot, and eventually one of those idiots it going to hit the environment with the cost of training another model of this caliber that does the exact same thing from scratch, just like Ling-1t and K2. So yes, I'm arguing it's also bad for the planet. This has implications that are hard to even dream of. And they are real, and they are not just a keyboard warrior on the other side of the world moaning and complaining for no reason.

Much like policy is a delicate dance that affects lives in real ways within workplaces and nonprofits, training data can have devastating and unforeseen results if it isn't good. It can be tough to see why at first, oftentimes. Licensing and the implications it has in the real world are similar. Of course, this is just one stranger's opinion on the internet. People will both agree and disagree with me.

Check out Richard Stallman's Ted talk or whatever if you want a primer. It's just a piece of honest advice in the best interest of a project I think is valuable to humanity.

Edit - I personally have convinced teams to modify licenses after internal review after bringing this to their attention, most recently, KAT-code. Go ahead and check their discord logs if you don't believe me. Their team apparently agreed enough with me after an internal review that even their 72b monster followed suit. MiniCPM before that, which happened in a hugging face discussion. Receipts are all online, real humans and real labs have aligned with this viewpoint. Not trying to namedrop to be cool or whatever, just to demonstrate that actual capable teams have looked at this argument of mine and found it compelling enough to put real job hours on it and make real, difficult decisions based on it.

Could it be that model weights are not software but data, and therefore, Apache 2.0 does not quite apply? Wouldn't you agree?

No and neither would a lawyer. Intellectual property is just data. I'd get sued from orbit if I started selling copies of LLama model weights, model weights/data or not. Plus, just because a system is complex and hidden behind matrix multiplication/coded differently than a human would, doesn't mean it isn't a system. There are dozens more reasons I disagree with this and I could go on. Sorry for the million edits, but I just also feel like it's important to note that, like you noted in asking if I agree, I'm just one fallible human being. Just because I disagree doesn't mean I'm right. Find the best path for your team based on your own intelligence. Cheers.

Yet another edit - Your arguments make a lot of sense through logic. My main argument is that this world doesn't subscribe to logic and so the implications of our actions don't either.

I don't quite see how your point is supported. If you explained how, the chances of us changing our position would be higher.
It also doesn't seem we are wrong about CC licensing for weights.
https://lfaidata.foundation/blog/2025/07/22/simplifying-ai-model-licensing-with-openmdw/

I am not particularly invested in this either way, and your position makes sense. I will talk with my colleagues when we get to it. Thanks for bringing this up.

Sure! As stated:

1. Environmental, as people will retrain for permissive license
2. Use case blocking for organizations and individuals that bar use of CC-by-4.0 tech in their stacks
3. Social, as the previous two points compound to result in complications in your initiative, which you earlier commented was "supporting underrepresented European languages." When accessibility is limited, complications arise, even if the limitations to accessibility are ONLY perceived by the user, as the user is the ones who brings the implications of these actions to reality.

A few other points:

OpenMDW-1.0 is a truly open license that is more copyleft than most. When I said "There is a reason more copyleft licenses than even those exist", this is exactly what I was talking about. It is radical in that it pushes for even the DATASETS used to train models be released clearly ALONGSIDE the weights, and also licensed in the way I'm advocating for, as well as including minor copyleft restrictions that ensure labs don't apply restrictions overtop of the license. This article you've posted is unfortunately (or fortunately for humanity, depending on your perspective) a HUGE self-slam to the point you're making; It's identifying that Apache-2.0 or MIT on a model weight is ACTUALLY NOT ENOUGH and argues that you should be going even FURTHER to copyleft/free software by ensuring the supporting architecture that functionally allows the model to exist within actual use cases are staunchly protected under truly permissive licensing, for ALL of the parts that affect the release.

If you agree with the article, switch to it for licensing and then everything's awesome. In short, what they argue is that even what I was pushing for, like a permissive license on the model weights, is not enough. It's not enough, likely in their opinion too, because of all of the same reasons I said CC-by-4.0 looked bad to me personally, though of course to say the authors don't like CC-by is a full on assumption with no hard evidence. Hope that helps.

I still seem to struggle to understand why this is a problem for European companies working in Europe.
I am not sure if we are even reading the same thing: The article clearly states that model weights are data, contrary to what you claim. So it's a view not only held by us.
Besides, it's not quite our problem that someone chose to restrict their stack to a set of licenses that do not permit using something applicable to data.
Part of the problem is not quite European. In the EU, there are no patents for software. Therefore, this is a problem only for companies competing in the US. But let's be honest - a model designed to be performant in Albanian and Icelandic, thus sacrificing performance in English?

I didn't say anything about Europe, North America, or anywhere else. Your "Well its not MY FAULT so why should I worry if people are using the cars I MAKE to obliterate humans? They're being silly! Driving wrong! Not my problem!" is foolish and laughable. Your first sentence doesn't actually make much logical sense at all. Do you... understand why this isn't a problem for one part of the world? What are you trying to tell me and why? If it's the rest of the world... let it burn? Only focused on "supporting underrepresented European languages" if it... happens in Europe? Sorry, but if there are people who are in massively English language dominated lands, especially ones where they can be a little bigoted to minorities (sorry America), they need this model more than literally anyone. I hate and do not give a damn about companies. I spend time in frustrating conversations like this because I am deeply passionate about the ESL community. Individuals. Real people. End users.

I explicitly never claimed model weights weren't data, as obviously they are and I firmly believe that as someone who works 60 hours a week in machine learning. I only mentioned what implications data has and that data is absolutely subject to license. In fact, here is me agreeing with you and the article multiple times from earlier:

"Data or not."
"Intellectual property is just data."
"Model weight/Data or not."

I was using YOUR framework (the model weights are just data) for all of my arguments, every single time, and you're still acting defensive.

Beyond that, you've called my personal intelligence and professional skills into question multiple times during this discussion, without it at all being needed. This is poor science. I am taking the time out of my day to help you because I believed in your project. Chill out and stop making your entire team look like a bunch of aggressive teenagers. I have given you all of the arguments you have asked for exactly, and you have not interacted with one. I have said that all I am offering is an opinion, and that it is up to you to find the best path for your team multiple times. You could have just thanked me for the feedback and gone on with your day like a sensible representative of a team.

I'm not even sure what you're talking about anymore, and your reading comprehension here isn't on par with your desire to split hairs within this context. It matters not if this is because of emotions or language barriers. You're fundamentally incorrect about most the things you're saying now, and there's no point me wasting any more time. You clearly do not understand (nor want to understand) the article you've sent me and are cherry picking, and besides that point, you're still stuck on something we've both explicitly agreed upon. I have had better discussions with conspiracy theorists. Model weights are data. This is a fact. The way you're throwing it around demonstrates profound lack of understanding, though. I tried, but I'm definitely done with this conversation. Last, as a few English second language folks have let me into their lives, I've had the opportunity to see myself what havoc machine translation can do as an abstraction layer on top of a communication. If you are running this through a translation layer, it's likely you aren't even reading anything you're being told incorrectly and that you're just getting bad translations. If the article or this conversation is being translated, I want to give a sincere warning that translation systems are not advanced enough for this level of debate in English.

In the future, if someone says something you disagree with, you can politely thank them for their time and move on. Or just be a keyboard killer, I'm not your mom.

Have a nice day!