My guess is that the meta team is currently playing around with qwen trying to figure out if their llama model can compare. If not, they might postpone the release...
I don't agree with people hating Llama 4, it's very useful as MoE, you can build computer with low VRAM and still get some t/s. I am waiting for Llama 4.1 and expect much improved models!
its also more designed for industrial use cases. not so much for hobbyists.
High memory usage, but very low compute + high parameter count = very good for industrial uses
I remember the original llama 3 also having issues, particular with non-english prompts, but that was completely fixed in 3.1. I hope the same is true for llama 4.
Indeed. On live-bench, amongst non-reasoning open-weights models, Maverick is second after Deepseek v3. But it's smaller and faster, so it's kind of expected to be slightly worse.
Nah. For a company as big and resource laden as meta, this was a weak offering which clearly shows a break down in their management or strategic focus.
It’s mostly because the release was done very poorly. Trust matters.
Scout don’t seems have captured a lot of interest, mostly because of Gemma 27b that is easier to run and better/equal to it. But Maverick did. It’s seems to be quite good for older/ddr4 server build. It’s roughly similar to a dense 70b, but faster. (And we did not have any good 70-120b models for quite some time. Command-a did not really pushed the boundaries.)
there are people in limbo like me. 96gb v ram but not super performant (m2 max). Where gemma3 doesn't quite cut it with speed. While t/s is not big deal difference 20 vs 30t/s (although noticeable) but the PP is drastic, sadly I don't have numbers at hand. Can't run maverick so llama4 is the best for my usecases currently. It's also actually pretty good, in some cases I'd say it's better than gemma or mistral small (which btw. I find better than gemma, except for pure language skills)
Just wait until we have quantum computing models then there'll be schrodinger's clopen weights and never even having certainty whether they're open and local, closed and local, open and cloud, or closed and cloud since the closer you look for locality the less you know about open.
> because there are limitations on how you can use those weights.
There is a piece of text that claim that they have ownership on the weight and that they are giving you a license and you have to adhere to it. There's no legal basis for that at the moment as model weights aren't copyrighted material under any jurisdiction now.
This is just an attempt to claim a new kind of IP right, and shall be disregarded (and I mean that not only because you shouldn't care, but because you shall refuse to care to stop them from being able to convert that attempt into an actual IP law in the coming years).
Yet there really is legal basis that it is not copyrighted material under any jurisdiction?
In any case, I don't think I would characterize it very open when using it against their terms could result in a long legal battle—if Meta cares enough about it. Certainly it could be a dangerous endeavour for small businesses.
> Yet there really is legal basis that it is not copyrighted material under any jurisdiction?
Copyrighted material is a term that is legally defined, and the current definition excludes model weights (as much as it excludes compiled artifacts: you cannot take someone else's code, and claim intellectual property over the compiled binary just because you compiled it by yourself).
> In any case, I don't think I would characterize it very open when using it against their terms could result in a long legal battle—if Meta cares enough about it. Certainly it could be a dangerous endeavour for small businesses.
I really doubt so, this “license” only works because there's ambiguity and losing a trial would end up destroying the ambiguity they are building upon.
If you are in the US, you'd have very good reasons not to do this since thanks to the broken legal system they can sue you into bankruptcy, but if you are in any country with a sane justice system, you'd be fine. They aren't gonna sue anyone in the EU who uses their model in complete violation of their license even if it's publicly advertized.
One could argue that it is a compilation, though? And then also argue that _creative imagination_ was used when building configuring the system that converted that material into the resulting LLM; after all, we can see that the capabilities of these systems vary a lot, while we can hypothetize that it is not only about their training material or parameter count that make the difference. So there's something else in play as well.
Truth is nobody has tried these in court yet (all the way through). We'll see what the NY Times lawsuit against OpenAI brings: if OpenAI loses, then it could mean a lot of these models would become legally undistributable.
> One could argue that it is a compilation, though?
Good luck claiming IP on a compilation from pirated material. It has to do *something* else.
> And then also argue that creative imagination was used when building configuring the system that converted that material into the resulting LLM
You can try compiling GNU software with a handmade compiler of your own (surely writing a C compiler requires creative imagination too), then release it with a proprietary licence and see how it goes. I'm not going to bet on your side.
> Truth is nobody has tried these in court yet (all the way through). We'll see what the NY Times lawsuit against OpenAI brings: if OpenAI loses, then it could mean a lot of these models would become legally undistributable.
This is the other side of the equation though, Meta/OpenAI could win their trial with their “it's fair use” argument and it still wouldn't make the model itself copyrighted material.
These trials annoy them very much, because it's going to remove the ambiguity and they have a lot to lose, but they didn't chose to start it.
No way they *initiate* a trial on the other side. They are just betting for a *“fait accompli”* with their licensing claims: after long enough of a shared industry coutume of adhering to model licences, they would end up having a *de facto* legal value that the judge will abid to (unless the legislator itself codified it litteraly in the law).
That's why we collectively **must not** show any regard to these claims.
The big companies almost can't start a legal case, as they would almost certainly be asked to show their training data.
And there are some very good reasons that the big companies won't be able to show their training data for the next couple of years.
It starts to get very interesting if somebody has a big gplv3 code base, which is part of the training data and they ask a model about their own code base, but the in between model is not open source...
DeepSeek also makes this false claim on their web site. Truth is none of them are open source - not any from the big players. None of the datasets and training recipes for these models are released.
And I see so few on localllama who care. As long as they have Apache 2 or MIT I'm good. I don't have the compute to repeat what they did or the money. As long as I have the freedom to use it without restriction and modify it I'm happy, but I sympathize with those who want and can do more.
https://preview.redd.it/vna7kgrlcmxe1.png?width=1122&format=png&auto=webp&s=2ee6720d9c97ada65314c7dac06f9d9d99422a97
In light of today's events, this joke is getting funnier and funnier
Isn't all news from Llama 4 BREAKING news? Or should I say, broke?
I hope this new news is that they fixed Llama 4 Scout to clearly out perform Llama 3.3 70b.
Not sure what your question is. Why would it do what? At present in my experience Llama 4 Scout acts around a 40b model with occasional jumps above Llama 3.3 70b, but it is not enough for me to toss aside 70b. Why would it do that? If they continued to train Behemoth and did further model distillation on Llama 4 Scout off of it, it has the potential to increase in performance.
As I recall, and perhaps faultily, Llama 3.3 was distilled from 400b model with similar performance as a result. In theory, I would say Scout could be trained about the same amount of time as it has been... with the finished Behemoth model and easily out shine 70b... but I'm no researcher, just have "more than a feeling" when I see the quantization having no harm to the model above 4 bit.
why would it outperform 70b.
rule of thumb for MoEs is usually sqrt(params*active params) so scout was aiming at very fast 30b and it delivered as you've said. I doubt it would change a week or so later as drastically to 70b. And your comment says "fix", that would be major breakthrough not fix.
BREAKING!!
Meta "may potentially release" a flock of birds into the auditorium.
Meta "may potentially release" a 1B model that beats Gemini Pro.
Meta "may potentially release" warez of that one Wu-Tang album.
The feck outta here with "may potentially". Fekkin influencer nonsense 🙄.
Considering R2 is also likely releasing this week, Zuck has a very small window to get any traction at all. R2 sounds like an absolute *monster* in cost savings alone.
Memory like in chatgpt? Isn't that just a short txt file that gets attached to the thread? And the new one seems to be RAG on previous threads. Does not seem hard to implement if you really need it.
But this is a major event so possibility is they may release new features and maybe smaller version of reasoning models ? We dk but hope they release something good and not like previous
71 Comments
no_witty_username@reddit
carnyzzle@reddit
Accomplished_Stay337@reddit
jacek2023@reddit
Expensive-Apricot-25@reddit
StyMaar@reddit
TheRealGentlefox@reddit
Expensive-Apricot-25@reddit
anilozlu@reddit
Soft-Ad4690@reddit
Mobile_Tart_1016@reddit
lily_34@reddit
ThenExtension9196@reddit
Serprotease@reddit
kweglinski@reddit
pseudonerv@reddit
DarKresnik@reddit
ColorlessCrowfeet@reddit
Calcidiol@reddit
eras@reddit
StyMaar@reddit
eras@reddit
StyMaar@reddit
eras@reddit
StyMaar@reddit
Former-Ad-5757@reddit
ColorlessCrowfeet@reddit
silenceimpaired@reddit
DinoAmino@reddit
silenceimpaired@reddit
qnixsynapse@reddit
DarKresnik@reddit
ApprehensiveAd3629@reddit
MoffKalast@reddit
pmttyji@reddit
Naurim@reddit
silenceimpaired@reddit
kweglinski@reddit
silenceimpaired@reddit
kweglinski@reddit
__JockY__@reddit
jnk_str@reddit
jnk_str@reddit
Worldly_Expression43@reddit
OkActive3404@reddit
2TierKeir@reddit
silenceimpaired@reddit
Content-Degree-9477@reddit
Remote_Cap_@reddit
nullmove@reddit
Remote_Cap_@reddit
silenceimpaired@reddit
Trysem@reddit
Luston03@reddit
Ok-Recognition-3177@reddit
power97992@reddit
jakegh@reddit
swagonflyyyy@reddit
Barubiri@reddit
nullmove@reddit
Thomas-Lore@reddit
nullmove@reddit
ZABKA_TM@reddit
Illustrious-Lake2603@reddit
wonderfulnonsense@reddit
Independent-Wind4462@reddit (OP)
a_beautiful_rhind@reddit
Few_Painter_5588@reddit
sunomonodekani@reddit
sunshinecheung@reddit
Namra_7@reddit