IffYuan / Embodied-R1.5-8B-SFT and IffYuan / EmbodiedVLM-8B-SFT-0128-7000

[dekelcohen]

Hi,
Thanks for the great work.
I noticed you are working these days on more advanced models.

Should we try them as a replacement to the 3B model ?

Thanks,
Dekel

[pickxiguapi]

Hi Dekel,

Thanks a lot for your interest and support!

First, a quick update: Embodied-R1 training scripts and dataset are now fully open-sourced, and we have updated the project to the ICLR 2026 version. The updated arXiv version will also be released soon. We warmly welcome you to use and evaluate it.

• Project repository: https://github.com/pickxiguapi/Embodied-R1
• Training scripts: https://github.com/pickxiguapi/Embodied-R1/tree/main/scripts
• Dataset: https://huggingface.co/datasets/IffYuan/Embodied-R1-Dataset
• Model (3B v1): https://huggingface.co/IffYuan/Embodied-R1-3B-v1
• Current arXiv page: http://arxiv.org/abs/2508.13998

Regarding the 8B model you mentioned: that is related to Embodied-R1.5 and is not an official release yet. We uploaded part of the checkpoints and dataset early, but we have not finished the usage guidance/documentation yet. Embodied-R1.5 is a major upgrade over Embodied-R1: it significantly expands general embodied foundation capabilities and achieves strong
performance across comprehensive VLM and VLA benchmarks.

The full release (paper + models + datasets) is planned for March — please stay tuned.