MolmoMotion¶

Development line: project:molmomotion · thread molmomotion-model-availability
Last event: 2026-06-18 · 1 dated since 2026-06-18 · Researched: 2026-09-05 · confidence: high

What it is¶

MolmoMotion is a Molmo 2-based 4B model for researchers building robot planning or trajectory-guided video generation.

• 3D point paths: forecasts object-attached paths.
• Conditioning: conditions on RGB history, 2D/3D query points, and an action.
• Model variants: supplies AR and flow-matching variants.

The H3-F30 checkpoint needs three frames and predicts 30 future frames, about two seconds at 15 fps. Use it for 3D motion-forecasting experiments, not as a text-to-video model or a robot policy by itself.

Development line¶

• 2026-06-18 — MolmoMotion-4B-H3-F30 model resource recorded. The autoregressive 4B checkpoint takes three history frames and forecasts 30 future 3D positions per query point.

What changed¶

2026-06-18 — MolmoMotion-4B-H3-F30 became available: the autoregressive 4B checkpoint takes three history frames and forecasts 30 future 3D positions per query point.

How to use this¶

As of 2026-06-18, track MolmoMotion-4B-H3-F30 as a separately identifiable MolmoMotion model resource. Verify its model card before relying on its capabilities, license, or deployment requirements.

1. Install the official repository in a Python 3.11 environment with the visualization extra. — https://github.com/allenai/molmo-motion
2. Download and load allenai/MolmoMotion-4B-H3-F30 with MolmoMotionProcessor and MolmoMotion; run the internal model in bf16 on CUDA. — https://github.com/allenai/molmo-motion
3. Provide three RGB frames ordered from earliest to t0, 2D point coordinates at t0, 3D point history in the camera-at-t0 frame, an action description, and future horizon 30. — https://github.com/allenai/molmo-motion
4. Call predict_trajectory and consume future_3d as an 8-by-30-by-3 tensor of meter-scale camera-frame coordinates; render or pass the trajectories to a downstream system. — https://github.com/allenai/molmo-motion

Best practices¶

• Choose H3-F30 for ordinary video input; choose H1-F32 only when there is a single query keyframe. — https://huggingface.co/allenai/MolmoMotion-4B-H3-F30
• Treat trajectories as planning or generation guidance and fine-tune for a downstream robotics task rather than presenting the base forecaster as a deployable policy. — https://huggingface.co/allenai/MolmoMotion-4B-H3-F30
• Use the official quick-start input contract; query-point 3D history must be expressed in the camera frame at t0. — https://github.com/allenai/molmo-motion

Superseded by this¶

• Nothing marked obsolete yet.

Still unknown¶

• We found no dated first-party successor release after the 2026-06-18 checkpoint event in the consulted sources. The public repository documents H1-F32 and model families, but does not establish a later release date for them.

Sources¶

Agent brief¶

• Subject: project:molmomotion, thread molmomotion-model-availability, 1 dated events 2026-06-18 → 2026-06-18.
• Practical note: As of 2026-06-18, track MolmoMotion-4B-H3-F30 as a separately identifiable MolmoMotion model resource and verify its model card before relying on its capabilities, license, or deployment requirements.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.