MolmoMotion-4B-H1-F32

This 4B model predicts three-dimensional point trajectories from visual observations and an action instruction.

Research release

What the release provides

The checkpoint uses 1 history frame and predicts 32 future steps. Output coordinates use meters in the camera frame. Inputs include query points and their three-dimensional history.MolmoMotion-4B-H1-F32 official model card

How to use it well

Practical guidance. Keep camera coordinates, time spacing and units consistent. Use trajectory errors over the full prediction horizon rather than judging only the first predicted frame.

Specifications & access

MolmoMotion-4B-H1-F32 specifications
DetailValue and conditions
StatusResearch release
AccessDownloadable
Model identifiersallenai/MolmoMotion-4B-H1-F32