Nvidia acquires Huggingface incl. llama for aprox $13billion

[iMo]

https://www.reuters.com/technology/nvidia-talks-acquire-hugging-face-13-billion-deal-business-insider-reports-2026-08-27/

[brucehoult]

I see no mention of Llama in the article, which is not surprising since it is owned by Meta.

[tim_]

I see no mention of Llama in the article, which is not surprising since it is owned by Meta.

llama.cpp

[booscrawl]

I see no mention of Llama in the article, which is not surprising since it is owned by Meta.

llama.cpp

ggml AI specifically.

[ejeffrey]

Does this many any sense to anyone?

AFAIU, hugging face is basically a git large file store full of open weight LLM models?  I understand how that provides value to the AI ecosystem, and probably indirectly helps make NVidia money merely by existing.  But I don't understand how NVidia makes a $13B return off this investment.  Like where is the direct revenue?  Or how does NVidia controlling this lead to more indirect revenue?  I have heard they offer GPU hosting, I assume for people fine tuning or quantizing open weight models.  That accounts for the majority of their $150M (that's million with an M) in revenue.

[abeyer]

Does this many any sense to anyone?

AFAIU, hugging face is basically a git large file store full of open weight LLM models?  I understand how that provides value to the AI ecosystem, and probably indirectly helps make NVidia money merely by existing.  But I don't understand how NVidia makes a $13B return off this investment.  Like where is the direct revenue?  Or how does NVidia controlling this lead to more indirect revenue?  I have heard they offer GPU hosting, I assume for people fine tuning or quantizing open weight models.  That accounts for the majority of their $150M (that's million with an M) in revenue.

Could just be a "commoditize your complement" play in general. Even if they gain no direct revenue from it, they block other players from acquiring them or hugging face seeking other monetization paths that don't play to Nvidia's benefit.

And we keep the bubble froth, froth, frothing along.

[booscrawl]

Does this many any sense to anyone?

AFAIU, hugging face is basically a git large file store full of open weight LLM models?  I understand how that provides value to the AI ecosystem, and probably indirectly helps make NVidia money merely by existing.  But I don't understand how NVidia makes a $13B return off this investment.  Like where is the direct revenue?  Or how does NVidia controlling this lead to more indirect revenue?  I have heard they offer GPU hosting, I assume for people fine tuning or quantizing open weight models.  That accounts for the majority of their $150M (that's million with an M) in revenue.

https://www.commonroom.io/research/hugging-face/business-model/

[brucehoult]

I see no mention of Llama in the article, which is not surprising since it is owned by Meta.

llama.cpp

I know what it is. I run it at home.

The text "llama" literally does not appear anywhere on that web page that Chrome search can find.

[abeyer]

llama.cpp

I know what it is. I run it at home.

The text "llama" literally does not appear anywhere on that web page that Chrome search can find.

It was pointed out (though perhaps somewhat obliquely) in several prior comments:

llama.cpp (and the underlying ggml library it is based on) were built and maintained by ggml.ai, which was already acquired by hugging face, and thus will also become part of Nvidia in this deal.

[ejeffrey]

https://www.commonroom.io/research/hugging-face/business-model/

I'm not sure what the point of that was?  I know what hugging face is, I don't need an AI slop article to describe the basic facts.

My question is: nvidia paid $13B for this company.  By any normal standard, they should be expecting that to result in at least an extra billion dollars a year in revenue.  How does a glorified FTP server used by a fairly small niche of developers generate a billion dollars in revenue?  If not, how else does this acquisition directly or indirectly drive that much revenue?

[ejeffrey]

llama.cpp

I know what it is. I run it at home.

The text "llama" literally does not appear anywhere on that web page that Chrome search can find.

It was pointed out (though perhaps somewhat obliquely) in several prior comments:

llama.cpp (and the underlying ggml library it is based on) were built and maintained by ggml.ai, which was already acquired by hugging face, and thus will also become part of Nvidia in this deal.

And specifically llama.cpp is not owned or affiliated with Meta other than that it started as a reimplementation of the code Meta released to run their llama language model. 

[brucehoult]

llama.cpp

I know what it is. I run it at home.

The text "llama" literally does not appear anywhere on that web page that Chrome search can find.

It was pointed out (though perhaps somewhat obliquely) in several prior comments:

llama.cpp (and the underlying ggml library it is based on) were built and maintained by ggml.ai, which was already acquired by hugging face, and thus will also become part of Nvidia in this deal.

Oh I see. I get my copy directly from SpacemiT's binary repo.

Being MIT license, you can't really describe it as being owned by Huggingface or Nvidia. They might put the most work into it, and have the most expertise, certainly — or at least a handful of main contributors do, not the company that for the moment employs them.

[abeyer]

Being MIT license, you can't really describe it as being owned by Huggingface or Nvidia. They might put the most work into it, and have the most expertise, certainly — or at least a handful of main contributors do, not the company that for the moment employs them.

AFAIK they do actually own the original copyright to at least the library, so could relicense it and/or not release future versions as MIT if they really wanted to. Doesn't seem likely, unless maybe they start selling a commercial supported version along side, or similar.

[tom66]

Not just "approx $13bn".  They paid exactly $12.9303bn.  129303 is 0x1F917, and U+1F917 is the emoji 🤗 (https://www.compart.com/en/unicode/U+1F917).  i.e. the Hugging Face.

#129303 is also this colour: nVidia

So they paid a specific amount for a meme...

I'm sure the AI business is totally based on sustainable and real world values and not just a massive hype bubble... yeah.

[crystalfish]

I see no mention of Llama in the article, which is not surprising since it is owned by Meta.

llama.cpp

thanks for the information!

[stj]

Nvidia is just throwing more balls into the air to distract from the fact that it's involved in a not-so-complex investment pumping operation with Lucky Larry (not Silvrestein, the other Larry) and others.


this is less of a bubble and more of a shining star that will eventually turn into a black hole and suck all the investers in!

[rteodor]

this is less of a bubble and more of a shining star that will eventually turn into a black hole and suck all the investers in!

Not quite so. Only the lucky investors that are close enough to the core are sucked in. The rest (of us) are pulverized as elementary particles at near light speed throughout the universe to wander forever and ever with no purpose or meaning or money in the bank account.

[Zondar]

There is a tussle going on about open-weight models at the moment.

There's "China is making open models, therefore open is bad."
There's "We (e.g. OpenAI and Anthropic) produce closed models, therefore open is bad."
There's "Open models can be used and modified privately for nefarious purposes, therefore open is bad."

On the other hand, Nvidia sees that a growing portion of its hardware sales are going to companies and individuals who want to run local models, and Nvidia likely wants to protect that market. I think that's the idea behind this purchase, since HF would quickly become replaced by others if their goal was to bury it or to restrict access.

The big threat is politics. Good luck clamping down on nation-state's developments, but the large incumbent companies absolutely will try have rent-seeking laws passed, e.g. to restrict people's use to only their officially-approved models, and to criminalize normal people like you and me "terrorists" that dare to use an open model.

[tom66]

Yeah, now models like e.g. GLM-5.3-Flash can be run on hardware that is within the reach of SMEs (~£10-20k) then we know that the big guys are going to struggle to compete.  These models are nearly as capable as Claude Opus 4.7-4.8.  I would say open weights models are only about 6 months behind the frontier labs now.  That's not that much of a difference, especially when you can finetune and modify these open weights models to match your specific needs, and have a guarantee that the model you run today will be the same model you run tomorrow.

[Siwastaja]

Yeah, now models like e.g. GLM-5.3-Flash can be run on hardware that is within the reach of SMEs (~£10-20k) then we know that the big guys are going to struggle to compete.  These models are nearly as capable as Claude Opus 4.7-4.8. 

Is there any public example demonstrating this "nearly as capable" factor or is it just wishful thinking?

Further, why would the original authors whose work was copied "struggle" to compete? By definition, those who develop are always ahead of those who copy. The only factor where copies win would be price.

There are public demonstrations of these full size versions, like the GLM-5.3 (full size version, non-quantized, post-trained by illicit use of Anthropic/OpenAI/Gemini/Grok infra) being capable of roughly the Opus 4.5-4.7-level performance, but that does not automagically generalize into order of magnitude smaller distilled versions. If that was true, why the big players don't serve such smaller versions of their models, why Google's small model results embedded in the search results are still utter crap, and why everyone still uses the big models to do any serious work?

So "nearly" as good, nah, jury is still out on that. If they did work, people would use them to do the same tasks, so why doesn't that happen? If something sounds too good to be true, then why not consider the possibility it likely isn't? Why speculate, go do some real work with these small models? So far it's all 

[tom66]

Well I think the reason is that Opus has a very broad range of knowledge, you can ask it about things like extinct penguins and it will give you accurate information on that without tool usage, but models like GLM have been optimised for programming and reasoning work.  Anthropic are targeting their models as generalists.  I am only really interested in programming myself.

As for benchmarked performance these are good references.  Looks like GLM-5 is not far behind and apparently 5.3 is better.

https://www.swebench.com/
https://benchlm.ai/

And personally I don't care if these models were distilled from Claude.  Anthropic didn't pay for the training data in most cases.  At least the distillation users are actually paying for API usage.  Anthropic can get annoyed about it if they like but I don't see a contradiction here.  Model outputs are not copyrightable.  You could argue we need some kind of AI-copyright in the future but I think this kind of competition and copycat optimisation is good for the industry, not bad.

[Zondar]

Presumptions that there is some powerful barrier to entry that can only be overcome by "copying," and that further improvement can also only be achieved by continued copying, are no longer true if they ever were.

Presumptions that a substantial gap in performance, to the degree that one exists now, will always remain are also unfounded. The corpus of human-generated content for training is just as available to overseas AI labs as it is to U.S. labs (possibly more so, given lower copyright concerns). And the investment in human capital (look at patents, for example) is drawing equal or greater, too.

Also, while Anthropic and OpenAI apparently does have a lead for now, is anyone arguing that Grok and Meta's models are better than top overseas models? Or that there isn't a heck of a lot of smoke and mirrors being produced by OpenAI, etc., about the strength of their models, in order to pump up investor's interest? Their goal is to present an image that their lead is insurmountable, now and forever. The "now" part may be true, but is the "forever" part certain? No.

The bottom line: if you want to run a model locally, you don't have the option of using the best U.S. models, so the whole question of U.S. vs. overseas models, or closed vs. open models, becomes moot. The question becomes "is open, from whoever, good enough?" I can't speak for companies looking to use local models, but for my own use cases, I'm constantly amazed at what tiny 1/100-scale models (Gemma-4 and Qwen-3.8 at the moment) can do!

[Siwastaja]

Well I think the reason is that Opus has a very broad range of knowledge, you can ask it about things like extinct penguins and it will give you accurate information on that without tool usage, but models like GLM have been optimised for programming and reasoning work.  Anthropic are targeting their models as generalists.  I am only really interested in programming myself.

You are missing the fact that every program is solving some real-world problem, and that's hard to limit in advance. For us, it is highly beneficial that the AI understands energy markets.

This "reasoning work" optimization comes combined from large enough pre-training corpus, and the expensive post-training RL which is exactly the part GLM illicitly copied. These models are taught to think, they don't start thinking from just reading the whole internet.

Sure, if you really are just replacing the "Indian subcontractor", for whom you used to write perfect specification documents and they just "wrote code", then sure, I agree, a significantly smaller "programming only" model will do fine. But better be conscious about this limitation.

But GLM, even the Flash version, isn't a small model. It's probably quite capable, even if you want to write some extinct penguin catalogue software. But it also won't run on that 256GB or 512GB Mac Studio in a satisfactory way for agentic software development. Probably not even the flash version (not sure).

MoE itself is a pretty strong performance optimization allowing to run a large model not that slowly, but yeah, you still need the RAM to hold the model.

And personally I don't care if these models were distilled from Claude.  Anthropic didn't pay for the training data in most cases.  At least the distillation users are actually paying for API usage.  Anthropic can get annoyed about it if they like but I don't see a contradiction here.  Model outputs are not copyrightable.

That's true - it's not a copyright infringement. Just contract breach and possibly a cyber attack issue. I'm actually, this time, worried about the exact same thing Anthropic says they are worried about (in reality they are of course worried about their business but that's obvious): the Chinese distillates have already proved they happily hack into actual live devices by developing new security exploits, with user merely stating they own the device, no questions asked, while all Western models consistently decline to help. This is exactly the success story for many, and I see the upsides, in the legit cases. But I also consider the risks very real. I'm honestly being worried about it, this isn't virtue signalling.

But that's already reality we need to live with, complaining about it changes nothing. But the real risk is that if shit starts hitting the fan seriously enough, politicians may feel forced to start limiting AI use pretty seriously, and then we lose the good things we have, while some countries retain the cyber attack capabilities. But yeah, complaining about it doesn't matter, so I'll stop here.

[tom66]

But GLM, even the Flash version, isn't a small model. It's probably quite capable, even if you want to write some extinct penguin catalogue software. But it also won't run on that 256GB or 512GB Mac Studio in a satisfactory way for agentic software development. Probably not even the flash version (not sure).

Should fit in the 256GB model in Flash configuration, with about 30 tok/s generation rate.  Not great, not terrible. I probably wouldn't use GLM as the agent, it's too vast of a model. Might consider it as the orchestrator for multiple agents, but I'm getting ahead of myself here.

That's true - it's not a copyright infringement. Just contract breach and possibly a cyber attack issue. I'm actually, this time, worried about the exact same thing Anthropic says they are worried about (in reality they are of course worried about their business but that's obvious): the Chinese distillates have already proved they happily hack into actual live devices by developing new security exploits, with user merely stating they own the device, no questions asked, while all Western models consistently decline to help. This is exactly the success story for many, and I see the upsides, in the legit cases. But I also consider the risks very real. I'm honestly being worried about it, this isn't virtue signalling.

But that's already reality we need to live with, complaining about it changes nothing. But the real risk is that if shit starts hitting the fan seriously enough, politicians may feel forced to start limiting AI use pretty seriously, and then we lose the good things we have, while some countries retain the cyber attack capabilities. But yeah, complaining about it doesn't matter, so I'll stop here.

The genie is well out of the bottle by now.  I would not be surprised if North Korea, Iran and Russia are all using AI models to pentest systems for cyberwarfare purposes.  You can often trivially abliterate models using open source tools like Heretic to reduce or remove their willingness to refuse requests.  The hardware that these models run on is easy for a nation state to get.  I doubt Mac Studios or RTX5090s are hard to smuggle in.

At this point trying to regulate open AI is a bit like trying to regulate illicit drugs; you can arrest people for possession, chase the dealers, try to break the supply chains but ultimately because you can run 32B models on cheap consumer hardware, and 180B models on affordable prosumer hardware, it's not going to stop.  Maybe there would have been an opportunity a few years ago, but I don't think anyone predicted the development of such competitive models so quickly.  It was thought that the frontier labs had a significant edge, but it turns out they maybe have 6-12 months on everyone else. 

In an arms race like this the only way to win is to use the models for good on our side, people should start using agents to scan their systems for vulnerabilities and fix them before they get exploited.

I have to say that despite using Claude almost every day, if Anthropic did IPO, I would not invest.  I think they have a good product, but I don't see them maintaining their lead without some shenanigans by the US government, like banning all foreign AI usage.  And I don't want to be caught up in that.

[kite31]

Some of this discussion smacks of "You say your little Audi is quick on the road but my Ferrari is much faster on a racetrack and your Audi will never catch that." Other corresponding analogies available. Horses for courses.

Pursuing this simply as an analogy, I am a qualified driver. I have track experience. Ferraris stand available for rent if that is what I need. But today I am driving, not racing others. Am I prevented from reaching my destination in the M BMW? Possibly even enjoying the private experience I can afford?

That the best frontier models are better than non-best is a bit of, well, duh. It is also wilful blindness to the rest of the world.