Locally hosted Kimi K3 （8X B300s） produced the most convincing physics across three browser-based crash scenes. Better than GPT 5.6
Test was done on atomic【.】chat， a desktop app that runs LLMs locally.
Each model had to generate a self-contained HTML simulation， so the test mixed coding， 3D scene design， and physics-engine configuration.
Prompts： - A monster truck crushing a row of cars - Two cars jumping a canyon and colliding head-on mid-air - A giant anvil drop test flattening cars one by one
Outputs： - Kimi K3 （local）： 32.7K tokens， $0 - GPT 5.6： 19.2K tokens， $0.30 - Grok 4.5： 55.2K tokens， $0.45 - GLM 5.2： 66.8K tokens， $0.15
These are quite tough experiments， the model has to create the geometry， assign masses and constraints， sequence the impacts， update object state， manage the camera， and keep the whole scene running inside one self-contained file.
Kimi K3 handled that loop best， particularly when damage needed to persist after impact.