Hugging Face SmolVLA

2025-06-03 · Publisher : Hugging Face · Domain : Robotics

Hugging Face releases SmolVLA, a 450-million-parameter open VLA that runs on consumer hardware, trained only on datasets shared by the LeRobot community.

### Background On June 3, 2025, Hugging Face introduced SmolVLA, a compact open vision-language-action model. While most VLAs have billions of parameters and need data-center GPUs, SmolVLA targets labs, schools and hobbyists with ordinary hardware. ### Specifications - **Size**: 450 million parameters. - **Hardware**: runs on consumer hardware. - **Data**: pretrained only on open, compatibly licensed datasets shared by the community under the "lerobot" tag on Hugging Face; the team improved task annotations and standardized camera views. - **Results**: SmolVLA-450M outperforms much larger VLAs on several simulation and real-robot benchmarks (scores published by the company). - **Ecosystem**: integrated into LeRobot, Hugging Face's open-source robotics library; Hugging Face acquired Pollen Robotics in April 2025. ### Why it matters SmolVLA reflects a frugality trend: smaller models trained on community data make robot AI accessible beyond large labs. It extends Hugging Face's open robotics strategy, from LeRobot software to hardware (Reachy, HopeJR). ### Key takeaway A useful VLA doesn't need billions of parameters: SmolVLA opens generalist robotics to small teams.

Tags : VLA, Open Source, LeRobot, Hugging Face, Frugalité