NVIDIA has released Cosmos 3 Edge, a 4-billion-parameter open world model designed for edge deployment on memory-constrained devices like NVIDIA Jetson, RTX GPUs, and DGX systems. The model combines an autoregressive transformer tower for vision/text reasoning with a diffusion tower for video, audio, and action generation, sharing multimodal attention layers for a unified world representation. It operates at robot-control resolution (640×360), generates 32 actions per inference at 15 Hz on Jetson Thor, and ranks #1 on VANTAGE-Bench among 4B-parameter models. A policy variant post-trained on the DROID dataset for pick-and-place tasks is also released, along with training scripts and a 4-step distillation checkpoint for Cosmos 3 Super that reduces inference steps from 35–50 down to 4 for up to 25× speedup.
