💻The power of open-source: In conversation with Thomas Wolf // TechBBQ 2026

You're right. You're right.

Yeah, we started in 2016 where basically language was not working. Everything in AI that was working was video, image recognition, like this classification, this 10,000 class that that image model could recognize. And so the original idea was really a game, and that's why we have this very non-serious name that I think is a bit maybe strange in today's world. So we had this idea of making a modern version of Tamagotchi where you could send a selfie to your little virtual being and you could talk about the image with it. And text was so bad actually that we even had a comedian in-house writing dialogues for our chatbots because the quality of dialogue generation with this, this was LSTM at this time, not transformers, and the quality of the generation was so bad that we just couldn't use that. And so one thing we were doing a lot was playing with the new tools from research labs, and in particular some of these tools being GPT-1 that a lot of people don't remember anymore. I think GPT-1, which was trained on a mostly love and romance novel. So GPT-1, if you had a beginning of a prompt with two persons, they would always kind of fall in love in some way in the continuation. I really like this model, but for us it was a very good model for dialogue. So we fine-tuned it to generate dialogue, and that's where I kind of discovered basically the power of these pre-trained models that are trained on very large corpora and that you fine-tune or you try to prompt or adapt to your specific use case. And so I open-sourced one repository to be able to use this model very well, and this became quite viral in a way in this small AI community of this type. And that's when we decided that maybe we should pivot there. We have on one hand this game that's kind of hard to make work, we have this comedian writing dialogue but that doesn't seem scalable at all. On the other hand, we have this open-source library where there's a lot of traction in this AI world, which was still quite small but we were already seeing glimpses of hobbyists playing with it, like glimpses of a world of AI that could grow quite a lot. So we decided to switch in 2019 and to convince our earliest investor that the future was not gaming but was open source. So we had a little bit of pedagogy to do that, but it worked well, and to raise our C on open-source AI.

To the comedian? I mean, he was really not, I mean his dream goal was not to be in a tech company to be honest. He was really trying to pursue the late-night evening stage in New York. So at some point we said, yeah, maybe you should pursue your dream and we'll do our dream.

So it's very interesting because it grew and we saw the pattern change. So for instance, at the beginning our library was mostly used over the weekend. So people would download it over the weekend. They would do their hobby projects around this neat AI thing, right? They would play with that and then on Monday they would be back to their boring job and then nobody would use our library. And so it was really driven at first by this mix of researchers of course but also hobbyists. And then we saw the pattern of download increase and then in the week, people starting to use it more also at work. And we have a survey we sent in the first years where we were asking, do you use the library for fun, for work, for both? And we saw it moving progressively from fun to both to work. And now we have a flat pattern of usage of course, and we saw that also increasingly moving from more research proof of concepts, hobbies to more professional people, basically full companies adopting the framework because it's very easy to use any type of open-source model. And then with the growth of open-source models, we also saw the framework just becoming kind of a production-ready library over time.

There was a couple of inflection points. All of this has been quite exponential in a way. So it's more like the slope of an exponential or maybe the exponent of the exponential just increasing. I would say for us there was three main stages maybe. So the first early one that we saw this exponential take off, but it is a rather small community, and that kind of were the first numbers we saw to tell our investors there is something here. Then we saw a second point when we expanded from just pure text to other modalities. And basically for us one of these points was Stable Diffusion, the release of maybe the first large-scale open-source image generation model. And here we saw this move where people just started to think, oh, AI is useful for text but also for image and then sound. And more recently, even very recently, we saw the DeepSeek effect, where basically I call that the ChatGPT effect of open-source AI, which means a lot of people at this time discovered that open-source AI was something. I even have my parents saying, oh, we're talking about open-source AI, because before that people had this, they started to use AI but they had no idea there was also this idea you could download the model and you could control the model and you could have an open source. So this was maybe our last inflection point in January this year.

The next one, there's a couple of very interesting trends. So my daily job at Hugging Face is mostly to see what are the next trends and how could we push for an open-source community there. I think last year it was agents was the next trend and now it's obviously something that everyone uses, right? It's even hard to use an LLM without agents now, without any tool. I think this year there is two, three very exciting trends. I think one is robotics in AI because it's also starting to work and there will be something happening at the end of this year or next year around this. We are very active in this field as well. Another one I think is AI for science, so thinking AI helping scientists to do things, which is very exciting in many sub-areas. It can cover weather forecasting, protein prediction, material science, fusion. We're active on all these fields with collaborations with scientific teams. And the last one I think recently we started to show was all this world model where you have not only watching a movie but you can start controlling the movie. You can start walking inside a picture. You can start interacting with something that's photorealistic. So I saw yesterday of course Victor speaking here. So he's deep in this field. I think this is going to be something very big because you can invent a lot of new things that blur the line between a film and a game, which was already, there was already a lot of cross-pollination with games being made out of film, with even film like The Last of Us being made out of a game. But now the line is going to be very thin. You have a full entertainment and even going to advertising, controlled advertising. So there's a full area of opportunity that's starting here and I think we're only beginning to explore.

We have, I mean for us the very long-term north star, the mission for us is we think like everyone here I guess that AI is going to be really transformative. It's going to be everywhere in our life like the internet. And we think there is a future where most of this AI is kind of closed-source, private thing, and there is another future that looks more like the future of the internet we dreamed in the 90s where a lot of this is open source and kind of community-built, like Wikipedia type of thing. And our main goal is to try to nudge as much as we can. We're still just a small company. We're no Google, no Amazon, no Meta. But we try to nudge the community in the other direction where kind of AI is also common good. I think this is good for many, many reasons. It brings less concentration of power, and concentration of power is extremely accelerated right now in AI. There's really this capital, this minds, everyone tends to cluster in this place. So we try to fight this trend. So for all these coming revolutions, robotics, science, world models, we try to push and to find the actors or to build the things ourselves that will make it more open source. So usually it's a mix of all of these. So we try to find good partners, people who are ready to train models, who have the knowledge. We try also to open-source the good basic building blocks. So we do a lot of work open-sourcing datasets for pre-training models, open-sourcing mesh datasets. So basically everyone who thinks I want to train a model can start from a very high-quality building block to get state-of-the-art models. And we also do a lot of work in open science, which means a lot of the science behind creating this model is getting closed. But if you rewind back like five, six years ago, it was not the case. A lot of things were published in conferences, like Google was one of the top paper publishers, right? But it's not the case anymore at all. And the first paper by OpenAI, they were amazing, and the GPT-3 paper is like 120 pages of knowledge. This is really great. This doesn't happen a lot anymore, right? Surprisingly, it still happened a lot in China, which is I think very, very surprising. The papers out of China and they publish it in conferences are still very, very high quality knowledge sharing. But to try to compensate that, we publish ourselves, we try to do a lot of publication, we try to encourage the groups who still publish, like Allen AI in the US, there's still a couple of very great groups. And that's what I call open science. And I think it's a very important building block of this. You need datasets, you need knowledge, and then you need money. And this we can't really help. We can't give a million dollars to people. But you can raise money for that.

Yeah. So this started two years ago where basically I saw a couple of interesting breakthroughs in robotics out of Stanford and Berkeley, and basically a robot started to be able to tie knots, to fold clothes, to even cook, like if you have a pan just throw an egg in. And I was like, this is amazing, there is something happening here. And I saw a couple of startups just after GPT-3, a couple of startups being there, starting to do things like Figure, 1X, but all of these are closed-source startups. And then there is a very clear future where we have a lot of robots there but they're all closed source, so we have no access on how they work, we cannot tweak them, we cannot understand really how they work. So we thought we have to do something, we have to try to create an open-source robotics community. So we started with software. I was lucky that one of the leads of the Optimus robot at Tesla just called me one day and said, I'm reporting to Elon directly every Monday and I'm getting a little bit stressed, a little bit fed up of all the stress of this 30 minutes direct confrontation with Elon. I want to maybe move out to open source. I said, why don't you join Hugging Face and we start to explore what we can do in open-source robotics. So Remy joined and we started one project that we called Le Robot, and this project was a huge success. I mean over a couple of months we had several thousand contributors and we had a lot of what we wanted, which is people who are software developers interested in robotics but a little bit scared of how to start, and they decided to join this project and to start to play with these new techniques. So this Le Robot project has been growing very strongly over the past year. And so at some point we were like, okay, we have a software project here but why don't we explore also open-source hardware because a robot is obviously a software stack but also a hardware stack, right? And in this we were working with one of the very few open-source hardware companies in the world, which is Pollen Robotics. There is a couple of others, there's Keylet in the US, like there's a couple of others but it's very few, very few companies do open-source hardware. And they were exploring fundraising for their next Series A and I was like, why don't you just join us instead of raising? And so we decided to join forces and we released our first robots in July called Reachy Mini, which is this very, very cute desktop robot, very cheap because one of our goals here is to make this extremely accessible. There's also a world where you have a lot of robots but they're kind of an elite possession, you know, maybe like the early cars, like the rich people have robots and poor people don't. I don't really want this type of future. I want a future where everyone can have a robot. So we decided to try to do a $300 robot. We released it in July and we thought maybe we got a couple hundred people interested in that and this also exploded. We sold a million dollars of these robots in five days. So last month I was trying to find suppliers and struggling to see how we can, but we're going to ship this thing, but it's going to be quite a challenge. So right now we're discovering basically the challenge of doing hardware, but I'm very optimistic. I think we're going to do very cool stuff there.

So the thing is you can do that. I mean the main challenge right now is you can do that in China. So the first supplier for us will be a good supplier. We've already been working with them, we did some cheap robots before. We did one $100 robotic arm, that's actually a great success called SO-100. And this one is our second. We would love in the future to be able to also manufacture these in Europe or the US. So that's our next challenge. I won't solve this in the next month, but I'm trying to solve that as well and see if we can manufacture that in a couple of places basically. Yeah.

Yeah, we did quite a lot in our life. We did six acquisitions. We do very strategic acquisitions, a bit like Apple. We try to find the team that have really, that are very dense in skills that we don't have. So for instance, Pollen is an amazing team in terms of technical knowledge on how to mechanically, how to build the mechanical, the electronics part of a robot. That was kind of knowledge we didn't have just previously at the end of last year. Because I mean, we're talking about robots but the big thing about Hugging Face is hosting this petabytes of model data, right? We really host petabytes, we have huge bandwidth, we just passed two million models hosted on the Hub, two million public models hosted on the Hub last week. And so we need to host this petabytes of data. So at the end of last year we acquired a company called Zetup which optimizes, you know, some of these models are a little bit the same because they have been fine-tuned but some parts of them are still the same or some parts of these datasets are the same. So we can do a lot of very smart duplication and Zetup is one of the companies we acquired to work on that. So we've done a couple of very targeted acquisitions there and I think we'll keep doing that but they need to be, I mean first teams that don't dilute the high density of talent at Hugging Face. So we need to really have this very high density of talent and teams that can also integrate in our very strange culture, I would say.