SmolVLA-450M
Small, community-data VLA that runs on a consumer GPU or CPU, trained entirely on ~23k public LeRobot community episodes yet matching much larger VLAs on real low-cost arms.
- 78%
- 65K
- 30 Hz
- 23K
Benchmarks
Numbers as reported by the authors on the named evaluation suite.
| Evaluation | Suite | Score |
|---|---|---|
LIBERO average | LIBERO | 87% |
Meta-World average | Meta-World | 75% |
Real SO-100 tasks | SmolVLA eval | 78% |
Async inference speedup | SO-100 deployment | 30% |
Training
Compact VLA (SmolVLM-2 + flow-matching expert)
450M
4x A100 80GB
380 accelerator-hours
210 h
23K
Vision-token pruning, layer skipping in the VLM, asynchronous inference stack for cheap deployment.
Fine-tunes on a single consumer GPU in a few hours from ~50 episodes.
Linked datasets, robots & papers
Resolving related models, datasets and papers…