Physics LLM 🪀

A fine-tuned LFM2-350M predicts 2D rigid-body physics frame-by-frame, as text — no physics engine. Pick a scenario, edit the scene JSON to move/add/remove objects, then watch it roll out live. The model sees as much of the prior trajectory as fits its 8192-token context.

Six scenarios (pong, bowling, ramp_roll, angry_birds, hourglass, newtons_cradle) were never seen in training.