OpenVLA 0.5B is a compact 0.5B-parameter model trained following the settings of the original OpenVLA.
It is initialized from a smaller Prismatic VLM checkpoint—Qwen2.5 0.5B—from MiniVLA.

This Prismatic-compatible checkpoint is particularly useful if you wish to fully fine-tune OpenVLA using native PyTorch Fully Sharded Data Parallel (FSDP) with the Prismatic VLM training scripts.
We employ this pretrained model as the initialization for CronusVLA-0.5B, a multi-frame visual-language-action (VLA) model.

Please refer to the official repositories for detailed usage instructions:

These repositories include guidelines for loading, fine-tuning, and evaluating this checkpoint.

If you find this model useful, please cite our work: