# AI2 Releases MolmoMotion, a Language-Guided 3D Motion Forecasting Model

_Research · published 2026-06-17_

The Allen Institute for AI (AI2) has released MolmoMotion, a model that predicts future 3D point trajectories for objects in video given a frame, a set of query points, and a natural-language action description. Built on the Molmo 2 backbone, the model forecasts where marked points on an object will travel over the next few seconds in world space — a capability aimed at bridging perception and physical reasoning for downstream systems. AI2 reports it outperforms existing motion forecasting methods on their new benchmark.

Alongside the model weights, AI2 is openly releasing MolmoMotion-1M, a dataset of 1.16 million videos paired with 3D point trajectories and action descriptions, as well as PointMotionBench, a human-validated benchmark of 2,700 video clips for measuring object-centric 3D motion forecasting accuracy. The project targets two primary application areas: robotic planning, where anticipating object motion before contact is critical, and controllable video generation, where trajectory-conditioned synthesis can produce more physically plausible frames.

## Sources
- [Hugging Face Blog](https://huggingface.co/blog/allenai/molmomotion)
---
Canonical: https://genbuzz.news/posts/ai2-releases-molmomotion-a-language-guided-3d-motion-forecasting-model