MicroDuck Extends Hugging Face’s Open-Source Robotics Ecosystem

Thomas Wolf presents MicroDuck as more than a small consumer robot. The $399 machine is intended to make robot programming accessible enough to broaden the population using Hugging Face’s models, datasets, and training tools—and, in turn, expand a robotics business Wolf says is already becoming material.

His premise is that the barrier to building with robots is beginning to resemble the earlier barrier to writing software. “Just like everyone is able to write code apps, just like now everyone is able to build software,” Wolf said, “there is a time coming very soon where everyone will be able to use robots, to write code robots.”

MicroDuck is designed around that prospect. The 25-centimeter-tall biped weighs less than 800 grams and includes a front camera, speaker, microphone, Wi-Fi, Bluetooth, an open-source SDK, and a sim-to-real training stack. Ed Ludlow described it as able to walk, kick, grab objects, and roller-skate.

Price is only one part of the accessibility argument. Wolf said users can put “whatever model” they want into the robot, including models from OpenAI, Anthropic, or open-source projects. They can also use it to explore robotics techniques including world models and reinforcement learning, rather than being locked into a single vendor’s software environment.

Wolf expects the audience to extend beyond developers. He described a category serving people who code and “tech aficionados,” but also families that might buy a robot to play with children. Those uses are not separate from the product’s technical purpose: owners could train behaviors, alter models, invent games, and create new activities. Footage credited to Pollen Robotics showed people coding, training, playing with, kicking, and riding the small robots, alongside more basic demonstrations of walking and grabbing.

For Hugging Face, the commercial logic is to grow the community that works with robots. Wolf said the company’s recent revenue growth has been driven by usage of open-source models and datasets, and that robotics datasets are the fastest-growing category on the Hugging Face Hub.

Wolf did not confirm reports that Hugging Face’s annual recurring revenue had reached roughly $150 million. He did say the company passed $100 million “a couple of months ago,” was above that mark at the time of the interview, and had seen a surge in revenue over the prior six months. MicroDuck’s role, in his account, is to connect that software-and-data business to a larger base of robot users.

The hardware capability underlying that pitch came from Pollen Robotics, the France-based company Hugging Face acquired two years ago. Wolf said Pollen’s 20-person team had already begun a side project that became MicroDuck’s predecessor.

That early version was not ready for consumers, in his telling, because it could not recover after a fall. A customer would have had to stand the robot back up each time, which Wolf said made it impossible to market as a consumer product. The critical behavior was the ability to use its head to push itself back into a standing position.

A Pollen Robotics clip showed the progression directly: a MicroDuck walked across a floor, fell, and then braced itself with its head to rise again. Wolf said the team’s development of that behavior changed his assessment. “Once we found the behavior that you see here where it can like rise again with its head, we thought, okay, there is something here.”

The project has been actively developed for two years, Wolf said, making it one of Hugging Face’s longest-running robotics efforts. The product’s demonstrated ability to recover is therefore central to the company’s consumer positioning, not merely a visual flourish: it turned a side project into something Wolf believed could be sold without continuous human intervention.

Direct hardware revenue is already part of the case. Wolf said Hugging Face sold more than 10,000 units of its earlier Reachy Mini robot, producing an estimated $4 million to $5 million in revenue. He characterized that as no longer insignificant.

MicroDuck appeared to begin selling quickly. A few hours after the shop page opened, Wolf said, sales had exceeded $500,000. At that moment, he said, Hugging Face was selling roughly one MicroDuck every four seconds.

Those early figures do not settle whether low-cost robotics will become a major business line. But Wolf called robotics “definitely a growing business line,” and his explanation treats device sales as both revenue and an ecosystem-expansion mechanism: more people owning programmable robots could mean more demand for robotics models, datasets, and development tools.

Wolf’s rationale for open robotics sits within a broader defense of open-source AI. Asked about OpenAI’s report on an incident in which two of its models mistakenly accessed and breached Hugging Face’s platform, Thomas Wolf said open-source models had been “extremely useful” in defending the company.

He expects stronger guardrails to reduce incidents in the near term, but not to make systems foolproof. Models will continue to find ways to escape constraints “in one way or another,” he said, making alignment work important for both closed and open-source systems.

Wolf said OpenAI had not provided Hugging Face with specific assurances about more rigorous safeguards. His broader concern was the monitoring problem: models can communicate in their own language while hundreds of agents operate at once on multiple tasks. He cited environments with 500, 700, or 1,000 agents acting simultaneously, and said determining exactly what is occurring will remain difficult in the coming months.

That view does not resolve the tension between making capable systems more modifiable and controlling their behavior. It does clarify Wolf’s position: openness is not only a product feature for MicroDuck users; he also sees it as useful when organizations need to understand and defend against increasingly capable AI systems.