Stream a 2.78-trillion parameter AI model straight from your SSD

You can now run a 2.78-trillion-parameter AI model on a laptop with a fast SSD.

An open-source C engine called WASTE lets standard consumer hardware run Kimi K3, a massive 2.78-trillion parameter model that usually requires server-grade memory. Created by sqliteai, the dependency-free tool streams model weights directly off internal NVMe storage on the fly. On a 64 GB MacBook Pro, it outputs exact, unpruned results at about half a token per second.

Why it matters: Mixture-of-experts models only activate about 4% of their parameters per token, leaving the rest idle. WASTE keeps the core model trunk resident in RAM and streams only the required experts off the SSD in real time. It trades speed for hardware access, proving frontier-scale models can run locally with zero cloud invoices and complete data privacy.

Know this: You need a terabyte of fast internal NVMe storage for the 982 GiB converted model, plus at least 29 GB of RAM to start it up. If half a token per second is too slow, the same engine runs the smaller 48B Kimi-Linear model at 10.7 tokens per second from a 19 GB container.

Turns out you didn't need a million-dollar server cluster—just a fast laptop drive and clever C code.