The Meta twitter account said "We’re also making it available in private preview via API to select partners, and we hope to open-source future versions of the model."
This is actually something that everybody on Earth is owed literally without exception. If I steal your car and sell it, are you seriously not owed the money? What if I went through lots and lots of hard work to steal it and sell it? I’d stay you’re still entitled to every dollar.
While that statement is true whenever something free is involved on the internet, let's not forget what we're talking about here. I don't recall having been asked for my shit to end up in their training data, either, so what about their entitlement? Seems to me there's some owing going on, at the very least.
Also, it's just a fair statement, I also hope a lot of things, but we all know what that word is doing there. It's a "hey talk about us we're awesome for helping out *maybe haha but also maybe not tehe".
Also also, it's fucking meta. Can't say I'd be awfully bothered about hurting their black heart.
Oh thank god they're going to open-source it. They're not the best lab, especially now, but I feel like America needs at least ONE somewhat major open-source lab.
Gemma models are tiny. They're great but there are zero American labs trying to make frontier large open-source models. Think the size of GLM 5 or DeepSeek V3.2.
Arcee AI released 400B Trinity Large Thinking a few days ago, and Trinity Large Preview a while back. That's the size of Qwen 3.5 397B and GLM 4.5/4.7 and Llama 3.1 405B. Not small, close-ish to GLM 5 and DS V3.2
And not very good from first tests I ran. It mixed up languages, wrote dialogue in one, story in another, and used my location data for story setting for no reason.
I mean, it's clear they don't want to support languages other than English.
Grok himself also used to mix English with another language. But Meta AI has had a supposed AVM for over a year now, and it's only available in English.
Soon when these big labs no longer release open weight models for free and we need to rely on the community models, we'll be back to models that does that.
If you're happy with them then I guess, but when that happens they will likely be the most capable local models you'll ever have, and any new ecosystem that renders them useless (for example, chat models -> agentic models already happened in the past) will mean we're back to square one.
Their licensing was always on the edge of acceptable to me… but their models were pretty powerful. I’d probably stick with Qwen 3.5 and Gemma 4 unless they gave a better license or incredible leap in tech.
If you access or use Llama 4, you agree to this Acceptable Use Policy (“**Policy**”). The most recent copy of this policy can be found at [https://www.llama.com/llama4/use-policy](https://www.llama.com/llama4/use-policy).
Also, I think there were Geographic Restrictions for some.
There was enough there they could rug pull if they were creative enough. Unlike Apache or MIT licenses. That said, they were far better than older Cohere licenses... newer license switched to Apache so that's nice.
(via WSJ) "In a departure from its previous models, which were open-source, Muse Spark is a closed model that will power Meta’s AI chatbot and AI features within it."
"the model is still underperforming on coding, so I would expect that to be a domain where they double down in the future."
The Muse Spark on meta.ai wrote a story for me mixing up two languages. I asked it in English, so it wrote the story in English but somehow put Polish dialogues into that and used my location in the story which was absolutely bonkers. There is no report button so I just downvoted it, but I have not seen a model fail like that since llama 2.
I was about to gift them one of my trickiest prompts as a goodwill gesture, a little homage to the Llama 3 days, but alas... you have to sign up. Hard pass, sorry, heh
From benchmarks, it looks to be a strong multimodal (only behind gemini). Its coding and reasoning abilities are behind OpenAI and anthropic.
A competitor entering with a strong model is a nice thing for us. Meta has one of the largest compute stack and large user base. I expect we will see prices from them only google will be able to match.
better than llama4, but this being a closed weight and falling behind all the other closed weights after spending billions on their superintelligence group - is not great
PERSONAL superintelligence - owned and operated by a CORPORATION. Come back when it can run local. Until then I don’t care how polite its personality is if it can’t be owned and operated by me.
drooolingidiot@reddit
The Meta twitter account said "We’re also making it available in private preview via API to select partners, and we hope to open-source future versions of the model."
Ok_Mammoth589@reddit
Yeah it's super easy to hope for something. I hope to win the lottery without playing it.
AnticitizenPrime@reddit
You're owed nothing.
mllv1@reddit
This is actually something that everybody on Earth is owed literally without exception. If I steal your car and sell it, are you seriously not owed the money? What if I went through lots and lots of hard work to steal it and sell it? I’d stay you’re still entitled to every dollar.
dhamaniasad@reddit
These models are trained on all of our data, aren't they?
ImpressiveSuperfluit@reddit
While that statement is true whenever something free is involved on the internet, let's not forget what we're talking about here. I don't recall having been asked for my shit to end up in their training data, either, so what about their entitlement? Seems to me there's some owing going on, at the very least.
Also, it's just a fair statement, I also hope a lot of things, but we all know what that word is doing there. It's a "hey talk about us we're awesome for helping out *maybe haha but also maybe not tehe".
Also also, it's fucking meta. Can't say I'd be awfully bothered about hurting their black heart.
russoue@reddit
Others are not even hoping :)
KaroYadgar@reddit
Oh thank god they're going to open-source it. They're not the best lab, especially now, but I feel like America needs at least ONE somewhat major open-source lab.
r15km4tr1x@reddit
Gemma?
KaroYadgar@reddit
Gemma models are tiny. They're great but there are zero American labs trying to make frontier large open-source models. Think the size of GLM 5 or DeepSeek V3.2.
FullOf_Bad_Ideas@reddit
Arcee AI released 400B Trinity Large Thinking a few days ago, and Trinity Large Preview a while back. That's the size of Qwen 3.5 397B and GLM 4.5/4.7 and Llama 3.1 405B. Not small, close-ish to GLM 5 and DS V3.2
zkstx@reddit
I would argue Acree can count as "trying"
r15km4tr1x@reddit
My interpretation of the release was that they created a small model for now they are scaling up but never said to what size would be open.
KaroYadgar@reddit
This could be true. We'd just have to wait and see I suppose.
r15km4tr1x@reddit
Exactly, and if they did get there, wouldn’t take much for Google to do it.
Belnak@reddit
Nvidia Nemotron, Arcee Trininty
Nghgminhtri@reddit
how about Mistral?
Belnak@reddit
French
sammoga123@reddit
There are about four test models. Among them are Avocato, and even one called Leviathan.
ApexDigitalHQ@reddit
Not open and no size??
Thomas-Lore@reddit
And not very good from first tests I ran. It mixed up languages, wrote dialogue in one, story in another, and used my location data for story setting for no reason.
sammoga123@reddit
The fact that Meta still doesn't have absolutely all the features beyond English and USA already tells you that things are going badly.
EbbNorth7735@reddit
Sounds more like bugs and an inference engine problem..
Paerrin@reddit
Why would they do work when they can just collect money and data?
sammoga123@reddit
I mean, it's clear they don't want to support languages other than English.
Grok himself also used to mix English with another language. But Meta AI has had a supposed AVM for over a year now, and it's only available in English.
LoSboccacc@reddit
oof, gemma 4 2b does that
po_stulate@reddit
Soon when these big labs no longer release open weight models for free and we need to rely on the community models, we'll be back to models that does that.
nabagaca@reddit
But we already have open models that don't do that, they can't take the models we already have weights for away
po_stulate@reddit
If you're happy with them then I guess, but when that happens they will likely be the most capable local models you'll ever have, and any new ecosystem that renders them useless (for example, chat models -> agentic models already happened in the past) will mean we're back to square one.
0xFatWhiteMan@reddit
lmao wtf
Faktafabriken@reddit
It looks like their idea is to offer ”personalised” experiences, and that they will collect your data to personalise your experience.
But we all know that meta is also selling data. My guess is that the user-data is what they will try to monetise. Might work…
SlaveZelda@reddit
Where did you run the tests? Meta App or an actual API?
Inflation_Artistic@reddit
Their website, meta(.)ai
ApexDigitalHQ@reddit
Thanks for the insight! I haven't played around with it yet but this is good to know.
MrRandom04@reddit
Huh, Meta finally got their lab back together. Shame they're most likely going to be private now.
silenceimpaired@reddit
Their licensing was always on the edge of acceptable to me… but their models were pretty powerful. I’d probably stick with Qwen 3.5 and Gemma 4 unless they gave a better license or incredible leap in tech.
CasulaScience@reddit
Their licence was literally use this for whatever you want as long as you're not Google or Apple
silenceimpaired@reddit
Not true.
https://ollama.com/library/llama4:scout/blobs/24ca191a372b#:\~:text=If%20you%20access%20or%20use,Llama%204%20safely%20and%20responsibly.
If you access or use Llama 4, you agree to this Acceptable Use Policy (“**Policy**”). The most recent copy of this policy can be found at [https://www.llama.com/llama4/use-policy](https://www.llama.com/llama4/use-policy).
Also, I think there were Geographic Restrictions for some.
There was enough there they could rug pull if they were creative enough. Unlike Apache or MIT licenses. That said, they were far better than older Cohere licenses... newer license switched to Apache so that's nice.
a_beautiful_rhind@reddit
As long as I have the weights they can write whatever they want in their text file.
Borkato@reddit
Lmao right?!
robberviet@reddit
Not open, not api, no tools around it, nothing. How can anyone try it? Oh just corp. Nevermind then.
JsThiago5@reddit
I dont know how much time it will be online but it created the hardest snake game I ever saw. You control the snake on the same plane using arrows and use w and s to change the 3D deep
https://embed.fbsbx.com/playables/view/4262558997332100/?ext=1783468980&hash=Q92gDAEZAZs2ixtCOefLZxczAQ_D
Eyelbee@reddit
Not open, boo
iDoAiStuffFr@reddit
quite close to sota as in quite behind
andy2na@reddit
look at the numbers again, they just highlighted their column, but most of the scores are not the best, see this for real benchmark comparison: https://www.reddit.com/r/LocalLLaMA/comments/1sfy877/meta_new_model_real_table_first_pic_vs_the_one/
iDoAiStuffFr@reddit
it's fooking shite m8 just like all meta models. alex wank my ass
Separate-Forever-447@reddit
(via WSJ) "In a departure from its previous models, which were open-source, Muse Spark is a closed model that will power Meta’s AI chatbot and AI features within it."
"the model is still underperforming on coding, so I would expect that to be a domain where they double down in the future."
...ok, then. Carry on.
tarruda@reddit
Benchmaxxed: https://x.com/fchollet/status/2042004767585751284
urekmazino_0@reddit
Meta AI engineer here - Meta is working biggggg with OpenClaw, our team recently hired 1000+ people for OpenClaw trajectory annotation.
westsunset@reddit
Are there other models in the family? Can you say the approximate model size?
sammoga123@reddit
That's why Meta bought Manus, right?
FullstackSensei@reddit
Not sure how I feel about that. But then again, I'm not a fan of Openclaw...
Thomas-Lore@reddit
The Muse Spark on meta.ai wrote a story for me mixing up two languages. I asked it in English, so it wrote the story in English but somehow put Polish dialogues into that and used my location in the story which was absolutely bonkers. There is no report button so I just downvoted it, but I have not seen a model fail like that since llama 2.
llama-impersonator@reddit
pointlessly chasing the hype, shocker
Thomas-Lore@reddit
It's not just hype, jesus this sub is stupid. :/
IrisColt@reddit
I was about to gift them one of my trickiest prompts as a goodwill gesture, a little homage to the Llama 3 days, but alas... you have to sign up. Hard pass, sorry, heh
LoveMind_AI@reddit
LOL. The website doesn't even work. :( haha
qwen_next_gguf_when@reddit
Mush Cock.
Charuru@reddit
RIP llama and open source
markingup@reddit
I thinks its actually pretty good tbh
Ok_Mammoth589@reddit
It's not even open weights...
TheDuhhh@reddit
From benchmarks, it looks to be a strong multimodal (only behind gemini). Its coding and reasoning abilities are behind OpenAI and anthropic.
A competitor entering with a strong model is a nice thing for us. Meta has one of the largest compute stack and large user base. I expect we will see prices from them only google will be able to match.
jacek2023@reddit
but no local (yet?) and I don't see the size
TheRealMasonMac@reddit
I think they said they would keep their largest models closed.
jacek2023@reddit
is this the largest one?
gavinderulo124K@reddit
No. They said their current approach looks to be a viable way of scaling up.
Hans-Wermhatt@reddit
Yeah, based on the results it doesn't seem like a smaller weight will come close to gemma or qwen benchmarks, but I'm excited for the release.
Dany0@reddit
Safetymaxxed means it'll perform below expectations. Also no announcement of even open weights. Wake me up wen gguf
CaptainAnonymous92@reddit
Wake me up wen you gguf gguf
Hefty_Wolverine_553@reddit
Benchmarks are pretty amazing if true, but doesn't seem like they're going to open source this one.
andy2na@reddit
look at the numbers again, they just highlighted their column, but most of the scores are not the best, see this for real benchmark comparison: https://www.reddit.com/r/LocalLLaMA/comments/1sfy877/meta_new_model_real_table_first_pic_vs_the_one/
Hefty_Wolverine_553@reddit
I know, but it's obviously a huge step up from whatever the llama 4 fiasco was.
andy2na@reddit
better than llama4, but this being a closed weight and falling behind all the other closed weights after spending billions on their superintelligence group - is not great
BIGPOTHEAD@reddit
Don't trust the Zuck
llama-impersonator@reddit
may meta get wanged to death
DrPaisa@reddit
I can't wait till meta tries to get a market share and hands out free quota gonna spam it like mad
silenceimpaired@reddit
PERSONAL superintelligence - owned and operated by a CORPORATION. Come back when it can run local. Until then I don’t care how polite its personality is if it can’t be owned and operated by me.