Hugging Face has launched an AI model for robotics called SmolVLA. This open model aims to democratize access to vision-language-action (VLA) systems, promising enhanced performance in both virtual and real-world settings.
"SmolVLA aims to democratize access to vision-language-action [VLA] models and accelerate research toward generalist robotic agents," writes Hugging Face in a blog post.
At 450 million parameters, SmolVLA can run on a standard consumer GPU or even a MacBook, making it accessible for hobbyists. It’s built on community-shared datasets and is part of Hugging Face’s broader push into affordable robotics with the recent acquisition of Pollen Robotics and the launch of its LeRobot toolkit. 
The new model supports an "asynchronous inference stack," enabling faster robot response times in dynamic environments.
Users are already experimenting with SmolVLA. One Twitter user claimed to control a third-party robotic arm with it, stating:
🚀 SmolVLA — feels like a BERT moment for robotics 🤖
I tried it on the Koch Arm:
Inference on RTX 2050 (4GB), fine-tuned with just 31 demos, and matches/outperforms single-task baselines 🔥
Big thanks to @RemiCadene @danaubakirova @mustash97 @francesco__capu 🙌 pic.twitter.com/TiBkAZGwkM
— Xingdong Zuo (@XingdongZ) June 4, 2025
SmolVLA isn’t the only player in open robotics. Nvidia and startup K-Scale Labs are also pushing into this field, alongside other companies like Dyna Robotics and Jeff Bezos-backed Physical Intelligence.
