DM05

DM05 is the base checkpoint of DM0.5, Dexmal's open-world Vision-Language-Action foundation model for embodied intelligence. It uses a Gemma3 4B vision-language backbone with a 680M Action Expert to generate continuous robot actions, and is designed for natural-language manipulation, zero-shot generalization, efficient downstream fine-tuning, long-horizon historical context, robust policy behavior, and transfer across robot embodiments.

How to Use

This model is intended to be used with OpenDM. See the OpenDM README for installation, checkpoint download, inference, fine-tuning, and evaluation commands.

Citation

@misc{dm05,
    title  = {{DM0.5}: An Open-World Foundation Model for General-Purpose Embodied Intelligence},
    author = {{Dexmal Team}},
    month  = {July},
    year   = {2026},
    url    = {https://www.dexmal.com/blog/dm0.5/index_en.html}
}