SmolVLA

Architecture

Compact VLA, 450M params

Key facts

SmolVLA release (primary Hugging Face)

Hugging Face's June 3 2025 post introduces SmolVLA as a compact 450M-parameter open-source vision-language-action model that runs on consumer hardware, pretrained on compatibly licensed community datasets under the lerobot tag. The base checkpoint is lerobot/smolvla_base. The post says the VLM backbone is SmolVLM2 and the action expert is a flow-matching transformer of about 100M parameters. Treat size and hardware claims as Hugging Face-stated ([Hugging Face](https://huggingface.co/blog/smolvla))

Reported performance (primary Hugging Face)

The same post says SmolVLA-450M outperforms much larger VLAs and ACT on LIBERO, Meta-World, SO100, and SO101, and that asynchronous inference is about 30% faster (9.7s vs 13.75s) with 2x task throughput (19 vs 9 cubes) at similar success. These are author-reported benchmark results, not an independent audit ([Hugging Face](https://huggingface.co/blog/smolvla))

LeRobot deployment documentation

The LeRobot documentation provides SmolVLA installation, policy configuration, dataset preparation, and training guidance for robot learning workflows ([LeRobot SmolVLA documentation](https://huggingface.co/docs/lerobot/smolvla))

SmolVLA base checkpoint

Hugging Face publishes the lerobot/smolvla_base checkpoint as the base SmolVLA model for use with the LeRobot ecosystem ([SmolVLA base model](https://huggingface.co/lerobot/smolvla_base))

Compact open VLA

Hugging Face describes SmolVLA as a compact 450 million parameter open-source vision-language-action model for robotics that runs on consumer hardware ([SmolVLA](https://huggingface.co/blog/smolvla)).

Community-data pretraining

Hugging Face says SmolVLA is pretrained on publicly available community-shared robotics datasets under the LeRobot tag ([SmolVLA](https://huggingface.co/blog/smolvla)).

Architecture

Hugging Face says SmolVLA combines a SmolVLM2 vision-language backbone with a flow-matching transformer action expert ([SmolVLA](https://huggingface.co/blog/smolvla)).

Asynchronous inference

Hugging Face reports that SmolVLA's asynchronous inference setup delivers about 30 percent faster response and twice the task throughput in its reported evaluation ([SmolVLA](https://huggingface.co/blog/smolvla)).

LeRobot integration

Hugging Face documents SmolVLA through the LeRobot framework, including the SmolVLAPolicy class and a pretrained smolvla_base checkpoint ([SmolVLA documentation](https://huggingface.co/docs/lerobot/smolvla)).

Developed by (1)

• Hugging Face

Common questions

Sources (3)

1. https://huggingface.co/blog/smolvla ↗
2. https://huggingface.co/docs/lerobot/smolvla ↗
3. https://huggingface.co/lerobot/smolvla_base ↗

Related brains

• Genie Operator-1 (GO-1)5 models
• Aurora Driver4 models
• Tesla FSD-Bot4 models
• Kodiak Driver3 models
• Nuro Driver3 models
• Redwood AI3 models