DwarfStar Inference Engine Enables 100T Parameter MoE on 96GB VRAM

DwarfStar uses a predictive expert paging algorithm to run Mixture of Experts models with 100 trillion total parameters and 80 billion active parameters. The engine achieves extreme memory efficiency by streaming experts into 96GB of total VRAM using four consumer-grade 24GB GPUs.