I follow llama.cpp pretty closely as I use either llama.cpp itself or projects that depend on it all the time, and one thing that I don't think gets talked about is the sheer scale of community involvement. It seems like a logistical nightmare, but somehow thousands of different contributors are opening hundreds of PR's every week and getting them merged in to support various hardware or implement a new pattern or algorithm from a recent research paper. It's really quite awe inspiring for me to see, and think it is in no small part because of the leadership of ggerganov - so I'm happy to see that he is sticking around and plans to keep building this incredibly useful tool that has grown into a huge community at this point.
Here's why I think it's working so well, as someone who's participating in llama.cpp development and who has participated in many other open source projects:
1) Reviews are super fast. Sometimes I get the first review 5 minutes after submitting a PR. Today I had a PR merged with two reviews within 42 minutes. This is incredible work by the whole team. It keeps contributors motivated. The pace is intoxicating. On other projects, I've sometimes been waiting months for a review.
2) Georgi motivates people by giving them responsibility. There's no gatekeeping as with other projects. He happily delegates. You do good work? It's appreciated, and you get the freedom you need to make an impact. That feels awesome.
I've rarely become addicted to an open-source project so quickly.
I have been using a PR branch to run GLM 5.3 flash locally so I have been watching the dueling PRs develop to implement it and all the comments and reviews. You're right the scale of it is huge. The speed with which maintainers and community are responding is also super impressive.
I know there's a lot of discussion about AI overwhelming OS maintainers and I can see that. It does seem like llama.cpp is successfully riding that dragon right now though.
Until 1-2 months ago, AI-written code was not allowed in llama.cpp. That changed, but it's still not allowed to have AI write your posts for you (issues, PR descriptions, etc.). In my experience, banning AI-written prose leads to much better contributions than on projects where it's allowed.
Of course, llama.cpp also gets lots of AI slop PRs from first-time contributors, but from what I've seen, maintainers simply ignore the bad submissions and prioritize the good contributions. I think that's probably one of the best ways to handle the slop overwhelm.
Folks are concerned that nvidia won't support these efforts if it gets models running on competing hardware. Two responses:
- The projects started without HF/nvidia involvement and were massively successful BECAUSE folks want to run models on their own devices.
- With highly capable agents, "hard to implement" should be less of a barrier. The inference market should in some sense become more efficient, as agents should make it easier to transition between software and hardware solutions. Sure there might be less training data for integration platforms, but if we've learned anything in the past few weeks, it's that agents can be remarkably persistent.
This is like if the local pool owner wanted to sell the property, and nobody in the neighborhood made a competing bid. HuggingFace isn't a publicly-owned resource.
What we should be seeing is competing businesses trying to undercut Nvidia/HF and offer a better experience. China has alternatives to HF; it's just that in the US, Nvidia is the only business that can get off their ass to sponsor infrastructure like that. America's poor underdogs like Intel and Apple simply can't be expected to make investments into real competition anymore. Nor can our startups be expected to find market traction when incumbents like HF are so deeply ingrained in the ecosystem.
Fact is, the pool was already crap. The rest of the industry had close to a decade to formulate any type of response, and America's vanguard took that time to rest on their laurels and formally forfeit to Team Green. Nvidia's played the "Luigi Wins By Doing Absolutely Nothing" card for the entire datacenter market: https://youtu.be/m6PxRwgjzZw
As a silver lining, it seems like we're hitting a huge compute limit everywhere we go due to TSMC's reluctance to scale. If demand for intelligence continues to stay high, an opportunity opens up to challenge Nvidia and Nvidia run kernels/models.
if you were a company on an island threatened by a billion-peopled country and your primary support was a senile corrupt democracy, would you feel like you could just roll the dice on bringing more factories into the world?
They don't have much choice. ASML is the bottleneck, TSMC is only the current hegemon. Nothing prevents Samsung or Intel from eventually catching up to their EUVL techniques, and in some respects they already have.
Welp, that's it then. Soon HuggingFace will require an NVidia Developer account, an onerous license agreement that must be agreed to during registration and a bloated cli interface with loads of telemetry.
I only need the license agreement to download and use CUDA/CUDNN, nvidia-smi, DLSS and proprietary drivers. Nvidia's not Apple, they don't make you register with them for basic SDK access.
Here's the content of the tweet so you don't have to give twitter any more traffic:
Hugging Face has been acquired by NVIDIA
It is quite exciting to be a part of this journey! NVIDIA has been an active supporter of the llama.cpp project. For more than a year now, their engineers have actively contributed to the codebase, collaborated with the community and provisioned hardware for development and testing purposes. The local AI ecosystem has largely benefited from our joint efforts.
Going forward, llama.cpp/ggml will stick to its founding principles. One of the most important qualities of the project is to be hardware-agnostic. Therefore the development and support of all backends will continue to be done as usual - driven and shaped by the community. Open to everyone who is willing to participate. The existence of such an independent software platform is crucial for the rapid adoption of AI locally and for bringing it closer to the user.
Now, with such a significant partner as NVIDIA supporting us, in addition to the wider Hugging Face team, I believe that our long-term goals will be easier to achieve. We will remain focused on building one of the most exciting projects based on the most exciting technology of our lifetime and making it truly available to everyone in the most accessible and efficient way possible.
Translation: Gervanov's ggml.ai was acquired by Huggingface in Feb 2026, so he is now "excited about the journey" after the Huggingface acquisition by Nvidia.
Can we take this as an official statement that Nvidia supports local models?
Why would Nvidia increase GPU efficiency for local models? Surely they'll operate like athletes and only establish a new record from time to time when necessary.
> Can we take this as an official statement that Nvidia supports local models?
The first reply on the X thread is by Gerardo Delgado, who is "Sr. Director of Product - Local AI, Creators, Developers @NVIDIA" according to his profile. In his post, he says "My team's goal at NVIDIA is to grow Local AI."
Why would they not? The cost to do that is 0.000001% of whatever numbers they usually work with. The upside is that it creates a community of hobbyists, most of whom will us (consummer grade) Nvidia GPUs. Which are not a significant source of revenue for Nvidia anymore, but it’s nice to have a bit of insurance.
> Can we take this as an official statement that Nvidia supports local models?
Hedging against a data center bubble burst or at least massive pullback
Trivial for them to cut margin on desktop cards and clean up/prop up falling data center sales serving the hungry gamer, blockchain, local AI that's been sitting on its hands the last year or two
I follow llama.cpp pretty closely as I use either llama.cpp itself or projects that depend on it all the time, and one thing that I don't think gets talked about is the sheer scale of community involvement. It seems like a logistical nightmare, but somehow thousands of different contributors are opening hundreds of PR's every week and getting them merged in to support various hardware or implement a new pattern or algorithm from a recent research paper. It's really quite awe inspiring for me to see, and think it is in no small part because of the leadership of ggerganov - so I'm happy to see that he is sticking around and plans to keep building this incredibly useful tool that has grown into a huge community at this point.
Here's why I think it's working so well, as someone who's participating in llama.cpp development and who has participated in many other open source projects:
1) Reviews are super fast. Sometimes I get the first review 5 minutes after submitting a PR. Today I had a PR merged with two reviews within 42 minutes. This is incredible work by the whole team. It keeps contributors motivated. The pace is intoxicating. On other projects, I've sometimes been waiting months for a review.
2) Georgi motivates people by giving them responsibility. There's no gatekeeping as with other projects. He happily delegates. You do good work? It's appreciated, and you get the freedom you need to make an impact. That feels awesome.
I've rarely become addicted to an open-source project so quickly.
I have been using a PR branch to run GLM 5.3 flash locally so I have been watching the dueling PRs develop to implement it and all the comments and reviews. You're right the scale of it is huge. The speed with which maintainers and community are responding is also super impressive.
I know there's a lot of discussion about AI overwhelming OS maintainers and I can see that. It does seem like llama.cpp is successfully riding that dragon right now though.
Until 1-2 months ago, AI-written code was not allowed in llama.cpp. That changed, but it's still not allowed to have AI write your posts for you (issues, PR descriptions, etc.). In my experience, banning AI-written prose leads to much better contributions than on projects where it's allowed.
Of course, llama.cpp also gets lots of AI slop PRs from first-time contributors, but from what I've seen, maintainers simply ignore the bad submissions and prioritize the good contributions. I think that's probably one of the best ways to handle the slop overwhelm.
Folks are concerned that nvidia won't support these efforts if it gets models running on competing hardware. Two responses:
- The projects started without HF/nvidia involvement and were massively successful BECAUSE folks want to run models on their own devices.
- With highly capable agents, "hard to implement" should be less of a barrier. The inference market should in some sense become more efficient, as agents should make it easier to transition between software and hardware solutions. Sure there might be less training data for integration platforms, but if we've learned anything in the past few weeks, it's that agents can be remarkably persistent.
While true, enshittification should be mourned every time it's renewed until some real kind of block (regulatory changes to the industry) exists.
This is like a company buying your local public pool, and you trying to convince everyone that things won't slowly turn to crap.
Yeah, in a year or so everybody will be surprised what became of huggingface and no one could have seen this coming with all the promises made ...
This is like if the local pool owner wanted to sell the property, and nobody in the neighborhood made a competing bid. HuggingFace isn't a publicly-owned resource.
What we should be seeing is competing businesses trying to undercut Nvidia/HF and offer a better experience. China has alternatives to HF; it's just that in the US, Nvidia is the only business that can get off their ass to sponsor infrastructure like that. America's poor underdogs like Intel and Apple simply can't be expected to make investments into real competition anymore. Nor can our startups be expected to find market traction when incumbents like HF are so deeply ingrained in the ecosystem.
Fact is, the pool was already crap. The rest of the industry had close to a decade to formulate any type of response, and America's vanguard took that time to rest on their laurels and formally forfeit to Team Green. Nvidia's played the "Luigi Wins By Doing Absolutely Nothing" card for the entire datacenter market: https://youtu.be/m6PxRwgjzZw
As a silver lining, it seems like we're hitting a huge compute limit everywhere we go due to TSMC's reluctance to scale. If demand for intelligence continues to stay high, an opportunity opens up to challenge Nvidia and Nvidia run kernels/models.
if you were a company on an island threatened by a billion-peopled country and your primary support was a senile corrupt democracy, would you feel like you could just roll the dice on bringing more factories into the world?
They don't have much choice. ASML is the bottleneck, TSMC is only the current hegemon. Nothing prevents Samsung or Intel from eventually catching up to their EUVL techniques, and in some respects they already have.
Welp, that's it then. Soon HuggingFace will require an NVidia Developer account, an onerous license agreement that must be agreed to during registration and a bloated cli interface with loads of telemetry.
I only need the license agreement to download and use CUDA/CUDNN, nvidia-smi, DLSS and proprietary drivers. Nvidia's not Apple, they don't make you register with them for basic SDK access.
Why are we not seeing xcancel alternative links?
X threw a fit.
Here's the content of the tweet so you don't have to give twitter any more traffic:
Did I ever tell you that you're my hero? You're the wind beneith my wings
"On Monday 24th August at 8PM EST, we received at letter from X Corp. asking to cease and desist the service XCancel.
The service XCancel is stopped until further notice.
We are seeking legal advice and won't share more details for now.
Thank you for the trust you have put in these two years of XCancel."
The answer to this question is, also, "X threw a fit". They sent Cease & Desists to nitter and xcancel.
They got a C&D.
Translation: Gervanov's ggml.ai was acquired by Huggingface in Feb 2026, so he is now "excited about the journey" after the Huggingface acquisition by Nvidia.
Can we take this as an official statement that Nvidia supports local models?
Why would Nvidia increase GPU efficiency for local models? Surely they'll operate like athletes and only establish a new record from time to time when necessary.
> Can we take this as an official statement that Nvidia supports local models?
The first reply on the X thread is by Gerardo Delgado, who is "Sr. Director of Product - Local AI, Creators, Developers @NVIDIA" according to his profile. In his post, he says "My team's goal at NVIDIA is to grow Local AI."
Why would they not? The cost to do that is 0.000001% of whatever numbers they usually work with. The upside is that it creates a community of hobbyists, most of whom will us (consummer grade) Nvidia GPUs. Which are not a significant source of revenue for Nvidia anymore, but it’s nice to have a bit of insurance.
> Can we take this as an official statement that Nvidia supports local models?
Hedging against a data center bubble burst or at least massive pullback
Trivial for them to cut margin on desktop cards and clean up/prop up falling data center sales serving the hungry gamer, blockchain, local AI that's been sitting on its hands the last year or two
Nvidia is the PC ecosystem's best answer to Apple
> Nvidia is the PC ecosystem's best answer to Apple
Sorry… what does this mean? I thought most people considered Apple irrelevant because of its low market share.
Macs are great for running LLMs. Apple has been investing in their M-series chips and it's been paying off, including when it comes to AI.