SmolVLA

SmolVLA is a compact (450M) open-source Vision-Language-Action model for robotics. Trained on community data, it runs on consumer hardware & outperforms larger models. Released with code & recipes.

SmolVLA is a compact, open-source Vision-Language-Action (VLA) model designed for robotics. With just 450 million parameters, it runs efficiently on consumer hardware like a single GPU or even a MacBook while outperforming larger models. Trained on publicly available community datasets, SmolVLA supports asynchronous inference, enabling 30% faster response times and double the task throughput compared to synchronous methods. The model combines a Vision-Language Model (VLM) with a flow-matching action expert, optimized for real-time control. Key design choices—such as visual token reduction, layer skipping, and interleaved attention—enhance speed and robustness. SmolVLA excels in both simulation and real-world tasks, demonstrating strong generalization across diverse environments. By providing accessible training recipes and affordable hardware compatibility, SmolVLA aims to democratize robotics research and accelerate progress in generalist robotic agents.