SmolVLA
Home > Models > SmolVLA
Key Significance
- Extreme Efficiency: 450M parameters achieving equal LIBERO benchmark performance (87.3%) as 7x larger Pi0 (3.3B)
- Runs Anywhere: Works on MacBook, consumer GPU, even CPU - dramatically lowers barrier to VLA research
- Community Data Only: Trained on 487 LeRobot community datasets (10M frames), achieves SOTA with only public data
- Asynchronous Inference: Separates action prediction and execution for 30% faster response, 2x throughput
- Fast Training: Completes training in 4 hours (20K steps) on single A100
- Flow Matching Applied: SmolVLM-2 + Flow Matching Action Expert architecture
- Fully Reproducible: Anyone can reproduce using open-source model, code, and public datasets only
SmolVLA: Achieving Pi0-level performance with 450M parameters
Overview
SmolVLA is a lightweight VLA model developed by HuggingFace, showing equal or better performance than 10x larger models with just 450M parameters. Runs on consumer GPUs and MacBooks, trained on LeRobot community data.
| Item | Details |
|---|---|
| Published | June 2025 |
| Developer | HuggingFace |
| Parameters | 450M |
| Paper | arXiv:2506.01844 |
| Blog | huggingface.co/blog/smolvla |
| Model | lerobot/smolvla_base |
Architecture
SmolVLA = SmolVLM-2 (VLM) + Flow Matching Action Expert
+---------------------------------------------------------+
| SmolVLA Architecture |
+---------------------------------------------------------+
| Inputs: |
| +----------+ +----------+ +----------+ |
| | Multiple | | Robot | | Language | |
| | Camera | | State | | Instruct | |
| | Views | | | | | |
| +----+-----+ +----+-----+ +----+-----+ |
| | | | |
| +-------------+-------------+ |
| | |
| +------v------+ |
| | SmolVLM-2 | Compact VLM |
| | (Context) | |
| +------+------+ |
| | |
| +------v------+ |
| | Action | Flow Matching |
| | Expert | |
| +------+------+ |
| | |
| +------v------+ |
| | Action | |
| | Chunk | |
| +-------------+ |
+---------------------------------------------------------+Training Data
| Item | Details |
|---|---|
| Data | 10M frames |
| Source | 487 LeRobot community datasets |
| Episodes | <30K (1/10 of other VLAs) |
| Environments | Various from labs to homes |
Performance
Benchmark Results
LIBERO Benchmark:
| Model | Parameters | Success Rate |
|---|---|---|
| SmolVLA | 0.45B | 87.3% |
| Pi0 | 3.3B | ~87% |
-> Equal performance with 7x smaller model
Other Comparisons
- Superior to ACT (simulation and real environments)
- Validated on LIBERO, Meta-World, SO100, SO101
Key Advantages
| Feature | Description |
|---|---|
| Lightweight | 450M parameters (10x smaller) |
| Efficient | Runs on MacBook, CPU |
| Fast Training | 4 hours on single A100 (20K steps) |
| Async Inference | 30% faster response, 2x throughput |
| Open-Source | Complete release of model, code, data |
Asynchronous Inference
SmolVLA’s differentiating feature: Separating action prediction and execution
- Low-latency control
- Real-time operation even in resource-constrained environments
- 30% response speed improvement
- 2x task throughput
Hardware Requirements
| Environment | Possible |
|---|---|
| MacBook | O |
| Consumer GPU | O |
| CPU only | O |
| Training (single A100) | 4 hours |
Impact
Significance of SmolVLA:
- Accessibility: VLA research possible without expensive infrastructure
- Efficiency: SOTA performance with less data, smaller model
- Community: Contributes to LeRobot ecosystem development
- Reproducibility: Fully open-source, public datasets only