SmolVLA
Hugging FaceSmolVLA is Hugging Face's compact vision-language-action model, about 450 million parameters, trained on community-collected datasets from affordable robot arms. It takes multi-view images, robot state and an optional instruction and outputs continuous action chunks by flow matching, trains on a single GPU, and runs on consumer GPUs or CPUs through the LeRobot library.
Openness
5 high confidence- weights
- open
- data
- open(the paper releases the community training data)
- code
- open(training code in LeRobot)
- license
- Apache-2.0(weights and code)
Weights, training code and training data are all released, under Apache-2.0, so the model can be retrained from what is published.
- https://arxiv.org/abs/2506.01844 recorded 2026-09-26
Paper abstract: "a small, efficient, and community-driven VLA ... We ... release all code, pretrained models, and training data."
- https://arxiv.org/html/2506.01844 recorded 2026-09-27
Paper: pretrained on 481 public community datasets from Hugging Face; "SmolVLA is pretrained on public community datasets".
- https://huggingface.co/api/models/lerobot/smolvla_base recorded 2026-09-26
Hub API for lerobot/smolvla_base: license:apache-2.0, gated false, model.safetensors (450,046,176 parameters).
- https://huggingface.co/lerobot/smolvla_base/raw/main/README.md recorded 2026-09-26
Card: "Reference implementation: https://github.com/huggingface/lerobot".
- https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/huggingface%2Flerobot recorded 2026-09-26
Repository record for huggingface/lerobot, the code: license apache-2.0.
Adoption
2 medium confidenceHugging Face downloads of the base checkpoint. The benchmark fine-tunes published beside it are not declared here.
- https://huggingface.co/api/models/lerobot/smolvla_base recorded 2026-09-26
Hub API for lerobot/smolvla_base: 71,362 downloads in the trailing 30 days.
Capability
2 low confidenceSmolVLA is built for the low-cost arms its community data comes from, and its real-world results are on SO-100 and SO-101 arms. It learns from far less data than OpenVLA and does not claim control of robots outside that family, though its paper reports results close to models ten times its size.
- https://arxiv.org/abs/2506.01844 recorded 2026-09-26
Abstract: "Despite its compact size, SmolVLA achieves performance comparable to VLAs that are 10x larger"; evaluated "on a range of both simulated as well as real-world robotic benchmarks".
- https://arxiv.org/html/2506.01844 recorded 2026-09-27
Paper: "we selected a subset of 481 community datasets obtained from Hugging Face, filtered according to embodiment type, episode count, overall data quality, and frame coverage" (22.9K episodes, 10.6M frames); real-world datasets collected "using the SO-100 robot arm ... and 1 with SO-101 arm".
Verified 2026-09-26