WARP

WARP is an embeddable inference engine written in C with no third-party runtime dependencies, designed to run large-scale mixture-of-experts AI models like Kimi K3 (2.78 trillion parameters) on consumer hardware by streaming activated model weights directly from fast storage (NVMe) instead of needing all weights in RAM. It enables running state-of-the-art language and multimodal models locally on machines with limited memory.