Matrix-Game 3.0: Real-Time and Streaming Interactive World Model with Long-Horizon Memory

https://arxiv.org/abs/2604.08995

720p up to 40 fps with a 5B model on an 8+1 GPU setup; 2x14B scaled model; camera-aware memory retrieval by pose and field-of-view overlap; minute-long consistency; DMD distillation, quantization, VAE decoder pruning.

Open real-time interactive video world models