Run Kimi K3 using 29 GB of RAM at 0.50 tok/s
Points and comments are a snapshot, not live.
WASTE runs the 2.78-trillion-parameter Kimi K3 model on a 64 GB laptop by streaming expert weights from NVMe.
WASTE is an embeddable C inference engine with no runtime dependencies. It keeps the model trunk in memory and streams selected experts from disk, using remaining RAM as a bounded expert cache. The converted Kimi K3 model is 982 GB on disk; minimum RAM is 29 GB, with tested speeds of 0.45-0.62 tokens per second on a 64 GB MacBook Pro M5 Pro. Overlapping expert reads with arithmetic and running the next layer's router early on the current hidden state doubles the hit rate from 14% to 38% without reading extra bytes.
A smaller 48B model runs at 10.7 tok/s from a 19 GB container. The engine refuses to start below the RAM floor and stops taking memory before the OS starts paging, which caps useful throughput on 64 GB machines at about 46 GB budget.