Paper

Ostinato: Fast CPU-GPU MoE Inference for Local Agents via Resident Experts in Every Forward Pass · ByteTrail