All right. Thank you. And I'm very pleased to have with me Clément Delangue from Hugging Face. Clément, you go by Clem, right?
Clem. Clem is perfect. Yeah. I've been in the US for 15 years now. So I've heard any versions of the pronunciation of my name, as you know, as a fellow French person. So Clem is usually the easiest.
Yes. And I apologize, you know, this is the part of the day where you get stuck in front of two guys who speak with a French accent. So, Clem, I mean, I think everybody probably knows Hugging Face in the room, but can you give us a quick intro? Mostly I think the founding story is pretty interesting and how you got to where you are with the twists and turns.
Yeah. And you've had kind of like a front row seat to that as being one of our earliest angel investors and supporters.
We started really out of a passion, out of an excitement for AI now almost 10 years ago. At the time it wasn't really called AI, right? People were talking more about computer vision, about machine learning, about deep learning. It wasn't as hot and sexy as it is today. And we really started with something completely different than what we do now. We started with conversational AI, kind of like a ChatGPT before ChatGPT, which obviously wasn't working at all with the technologies of the time. And after three or four years, came out a model from Google called BERT, and one of my co-founders Thomas was like, okay, BERT is cool but it's in TensorFlow. I want to spend the weekend putting it to PyTorch for the community. And we did that. When we released it a few days later, it blew up online. And that's when we realized that we can not only be conversational AI but be more of a platform for people and companies to use open source models.
Cool. Could you give us a sense of the scale of the company and the user base, your place in the ecosystem?
Yeah, so we're now the biggest platform for AI builders to use open source models, open data sets. We have 15 million AI builders using the platform today. They're creating a new repository on Hugging Face — so a repository is a model or a data set — every 7 seconds now. And there's like 3 million open public models on the platform, over a million open data sets. So it became really the central place for people and companies to build their own models, especially when they don't want to use APIs.
Cool. And people can run the models also on your platform?
Yeah. Usually either using our compute or integrated with other platforms and other partners. We really have the open source philosophy that our platform is built in a way that it integrates with the whole ecosystem and also with your internal systems and internal stack. So if you want to use part of the Hugging Face platform, connect it to your internal stack or to someone else's stack, it's quite easy to do.
Cool. So I very distinctly remember the first time we met. I think it was at a data meetup in New York, in late 2019 or early 2020. And you made that super convincing pitch around the fact that language was solved and that absolutely everything would be disrupted by AI and AI was taking over. And to put that in perspective, it was before ChatGPT. It was even before GPT-3. I think it was after GPT-2.
So of course since then pretty much everything you said at the time came true. So what are you seeing now? What is going to come true in the future? Can you educate us a little bit? The billion dollar question, right, for the next steps.
What we're seeing on the Hugging Face platform, which is sometimes kind of like a sample of the future because we have the most expert and most nerdy AI builders using our platform, is a lot of traction and interest for everything except text now. We see a lot of usage on biology, chemistry, time series, robotics. So I think we see AI starting to permeate really all the other parts of the economy. I'm super excited about it because thinking about AI for biology and how AI can help on drug design, on DNA analysis. We released a model two weeks ago that allows you to analyze your DNA on device, on your laptop. So these are the things we're going to talk about much more in the next few months. The second thing we're seeing, which is a bit counterintuitive, is more and more smaller, more customized, specialized models. Everyone talks about the gigantic models behind APIs, but I think we're going to see in the next few months that sometimes using smaller, more specialized models is a better option.
Cool. I think we can come back to some of that. I think there's a lot of questions around how each of us should be thinking about those models, whether we should adopt the big ones, the small ones, build a smaller one, and what the path from here to there can be. So if we switch gears a little bit, you run one of the largest open source platforms basically. I mean there's GitHub for code and there's Hugging Face for AI, that's more or less the split. And there's a lot of look about what it means to be open source for AI. There's open weights, there's open code, there's open data sets. Is there a right way or wrong way to think about it? How do you like to think about it?
No, I like to think about it as a gradient. You have kind of like pure closed source, black box APIs where you send an input, you get an output. You don't really know anything that's happening in between. Could even be a human interacting in the middle. You have no idea. And you have the pure open source where you have open source data sets, open weights, the scripts to create the models, train the models, run the models being open. Most of the workloads happen in between. And the more we can push the gradient towards openness, the easier it gets for everyone to participate in the creation of AI and not just a few companies. The more it creates transparency for everyone to understand what's happening, why it's happening, and the more control you give to every company, every individual, every type of organization — not just big corporations but also little tech, nonprofits, academia of course. So this is a gradient and hopefully as a society, as an economy, the more we push towards open, hopefully the better the outcomes.
And you do push towards open, right? So you have a research team, you build models, you publish models, and you publish the whole thing — the data sets, the training, and everything else, right?
Yeah, yeah. We did for example something called SmolLM, which is very competitive on the Pareto curve of small language models, where we published not only the weights as you mentioned but also the data sets and the full training scripts so that anyone can basically retrain these models from scratch if they want to.
And what do you see the rest of the industry doing there? There's been a lot of debate over the past few years around the big labs not publishing anymore. Whether that's OpenAI, Anthropic, Google — word on the street is if it's published it means it's not interesting anymore, like the good stuff they keep for themselves for competitive reasons. Do you see that?
We see that. When we started Hugging Face, the whole field in the US was very open with open research, open source weights, and that's very much what propelled the sector in the US. If you think of ChatGPT, the T in ChatGPT is actually Transformers, which was an open source contribution by Google. So this kind of emulation and collaboration of the whole field, in my opinion, is what created the American leadership for the past few years. It kind of stopped or regressed with a lot of these organizations being more closed source, being more secret, focusing more on monetization than open sharing. We've seen the opposite in China where they've started to be more open, share more of the research, share more of the models, which led to this kind of weird point in history today where a lot of the best open source models are actually coming from China and getting a lot of usage in the US.
So we see a lot of companies including enterprises now using open weights models from China in the US. And on your platform, you see the most reuse and forking and downloads for Chinese models basically.
Yeah. End of last year, for the first time, the volume of downloads of Chinese open weight models has been higher than the volume of downloads for American open weights models.
And do you see a turn there? There's a few new labs, whether that's in Europe or in the US, that are at least planning to release open models.
Yeah, I hope so. We have an administration in the US now that is really in favor of open weights, open data sets. The AI Action Act from I think it was a year ago — the open data sets and open weights was the third priority of the White House guidelines. And you see some organizations doing more. OpenAI released some open weights a year ago. You have NVIDIA that is now the biggest contributor to open source in the US pushing the field. We announced actually today a collaboration with a company called AI2, which is a smaller AI lab sharing open weight models. So I think the US could get back to their roots and get back their leadership. It's going to take some time, but I think it's possible and it's something we should fight for and help because that would be a great thing for the US and for the world.
Yeah. And I think for everybody in the room who got to a point where they tried to customize their own models, a few years ago you would start from Llama. And today you mostly you're going to start from Qwen or one of...
Yeah. You see tons of examples like Cursor, right? They announced I think three months ago that they were post-training their own models. And a company like Cursor, a darling of Silicon Valley, funded in the US — you would think that they would use American open base, but they disclosed that they were actually basing it on Chinese open source. And I see not even dozens, hundreds of these examples. When I talk to people they usually don't disclose it because it might not become the most popular thing to disclose in the US. But the volume of usage of Chinese open weights in the US is really surprising, not only in companies but also in academia. Now if you go to Stanford, to Harvard, the only way to study the state-of-the-art in AI is to look at Chinese open weights, which is kind of weird, but that is where we are today.
And some labs, I know DeepSeek is publishing a lot also, so there's a lot of information that comes out there. Great. So this gets us to the idea of building with AI, not just consuming it. And I've heard you convey the message that it's important for all companies to be builders and consumers of AI.
Can you tell us a bit more on that? Like what does that mean? Can everyone build?
Yeah. So I think the way to approach AI and to look at AI is kind of like as the new way to build software. It's this analogy from Karpathy of software 2.0. And I think what we've learned with software is that if you're a company trying to play a role in technology, you have to build the capabilities to write software yourself. You can't outsource all your software to someone else. And I think the same goes for AI in the sense that if you're using only an API, sending your data, sending your queries to an API provider and actually not building anything yourself, long term, I really think you're going to be at a disadvantage. First, this provider is going to try to take advantage of you and milk you as much as possible with their margins, sometimes by building themselves competing products to what you actually build and pricing it to a level where it's impossible for you to compete. And second, you're going to end up with a model where it's way too costly, way too slow, and it's just going to be a struggle for you to actually build your competitive advantage.
Well, that's uplifting. So everybody should build, otherwise we'll be trapped forever. No, but seriously, I think you make a great point. I think what everybody needs to realize, it's an investment because mostly initially it's not going to pay off immediately.
It takes time to get those technologies to work. And by the time you customize the model, a new commercial one comes out, so it doesn't look like you've moved much. But you're in touch with the market and then you can learn and improve.
Basically similar to software, right? If you weren't writing software before, the transition to software wasn't always easy. At the beginning maybe people thought, okay, I'm not a software company, I don't need to write software myself. But they learned, they built the capabilities and progressively they saw the value in it. It's the same thing for AI in my opinion.
Right. So another thing you have a great viewpoint into is the various trends in adoption. And in particular, it's very interesting to see right now how various parts of the world react to the rise of AI. We've all seen those polls that say that AI is extremely unpopular in the US, even more unpopular in parts of Europe, actually quite popular in Brazil and very popular in China. What do you see in terms of the adoption on your platform? Where are people downloading the models the most and building the most?
In terms of builders, obviously the US has a great culture of building with AI, so that's where we see the most usage. Obviously a lot of usage in China, even if we're banned from China as an American company, but we still see a lot of usage from there. But we definitely also see this trend of people being scared about AI, having negative opinions about AI. To be honest, I think part of the problem comes from the AI leadership that sometimes talks and emphasizes too much the wrong issues. Don't talk enough about how AI can create jobs, how AI can solve new challenges, how AI can change the distribution of power and wealth. And instead tend to talk about catastrophic scenarios for AI, like Robocop, where AI is going to take over the world, how AI could destroy every single job that ever existed. And so that created this weird perception in the public.
But you see the builders are still fairly insulated from that.
Basically. And I even see the correlation meaning that if you give people more tools to build themselves, I think they're going to see AI for what it is, which is the new way of building software, a new tool that they can use to solve their problems. And I actually think that it's going to help on their perception. And we see the divide a little bit. In the US a lot of AI is closed source, bad perception of AI. In China a lot of it is open source and actually you have a better perception. It's obviously not only correlated to that, there are a lot of other factors, but I think it plays a role.
I see. So talking about Robocop scenarios, so you're getting into robotics too, but in a good way, right?
Yeah, not a Robocop kind of thing. It's not going to kill you, it's not a giant humanoid robot that is going to do a force movement and kick a kid in the stomach. I don't know if you've seen this video online.
I haven't seen that one, but it sounds pretty horrible.
Yeah. It kind of looks funny at the same time, but if I were the parents of the kid, I wouldn't be happy about it. I mean a year and a half ago, we started to see more and more people talking to us and being like, okay, I'm an AI builder, I'm building with robotics, but I don't really have the hardware to experiment with it, because it's very expensive. And so we acquired a company back in France called Pollen Robotics that had been doing robots for seven years when we acquired them. And now we built this little robot called Reachy Mini, which is super cute, that only costs $500 and it's kind of like an open source platform for robotics, meaning that you can program it yourself. It's fully open source. So a lot of people use it to experiment with robotics. And it opens a window into new tasks, new use cases, new capabilities. Like people put it on their kitchen counter to help them cook. Some people give it as an interface to their kid, without the kicking in the stomach, for them to interact with AI without having to look at the screen, and to learn a new language, new topic, things like that. So yeah, it's a fun extension of the use cases and it's an example of what I was saying, meaning that AI is really permeating from only being a text assistant to really now getting into robotics, into the real world, into industry, into biology, chemistry, and all these other domains.
And that's really exciting. We've in a way we live in science fiction now, we talk to the computer, the computer answers. And we live in a world where more and more people start experimenting with robotics. I mean when I was growing up it was really not like that. So despite all the doom and gloom that we can read, I think that part is pretty exciting.
My dream is that in the next few years you have a hundred million AI builders being able to really build for their own problems, for their own challenges. And I think that would be a really great world if we can enable and empower more people to be AI builders, because it means that we won't outsource that to just a couple of labs in Silicon Valley. I think the world where you only have a few companies in Silicon Valley building AI for the rest of us is kind of scary when it comes to concentration of power, concentration of wealth, concentration of capabilities. And so if you can change that with open source and make more people become AI builders, I think we'll be in a way better world than without it.
Thanks. Sounds like it's a popular idea. Have you thought of running for office maybe?
No, no, no. An AI would be better than me at running some countries.
Well, maybe in a few years, give it time. Okay, so you see everything that's changing in AI, things are changing so fast. I literally spend 45 minutes every day catching up with what happened since the last day. What do you see — are there big changes that you didn't anticipate like one or two years ago?
