A mixture-of-experts inference engine that keeps the model on disk and moves only what the next token needs.
The router decides which experts a token needs. Sleeper fetches those and leaves the rest where they are — so the memory ceiling is set by the resident window, not by the size of the model.
Mixture-of-experts models are mostly idle. Any given token activates a small fraction of the total parameters, but conventional runtimes still expect the whole set to be addressable in memory. That expectation is what puts frontier-scale models out of reach of consumer hardware — not the compute.
Sleeper streams experts on demand. It reads the routing decision, moves what that decision requires, and holds a fixed-size resident set. The engine has been validated running a 744B-parameter model on a single consumer GPU.
Expert streaming is an active area and other runtimes have shipped offload strategies that close part of the gap. Sleeper's numbers are published as a repository you can run rather than a chart you have to trust. If a benchmark here doesn't reproduce on your hardware, that's worth an email — it's the only kind of correction that's useful.
Benchmarks: github.com/paraxaQQ/sleeper
Everything else: ashton@actualintel.co