Physics LLM 🪀

Qwen3-0.6B-physics predicts 2D rigid-body physics as text — no physics engine. A Qwen3 body retrained on a fixed-point tokenizer (vocab 1024) emits a whole trajectory per forward pass; we re-seed it with its own last frame (sliding window) for long rollouts, decoded with n-gram speculative decoding (~3-4× faster). Pick a preset, drag the balls around or drop new ones on the canvas, then watch it roll forward live against a Pymunk ground truth.

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

✨ Featured demos (model handles these cleanly) — click to load: