Moonshot Dropped Kimi K3's Weights on July 26 and Three Consumer-Hardware Inference Engines Appeared Within a Week

GitHub·high signalXBlueskyLinkedInCopy link

WASTE (sqliteai, 1,366 stars, created 2026-07-28) runs the full 2.78-trillion-parameter Kimi K3 on a 64 GB MacBook Pro at 0.45 to 0.62 tokens/sec by keeping the 27.28 GB trunk resident and streaming selected experts from NVMe, with 3-bit residual vector quantization on experts and a 982 GB on-disk container versus 1.42 TB published weights. The README documents a counterintuitive failure mode: growing the expert cache from 17.32 GB to 29.32 GB raises hit rate from 36.2% to 41.3% while throughput collapses eightfold, because a cache hit becomes a page fault. Moonshot released K3's open weights on 2026-07-26 with official guidance calling for 64+ accelerators, so this entire class of repo is a five-day-old reaction to that drop.