WASTE, a new open source inference engine, ran a mixture-of-experts model built with 2.78 trillion parameters entirely on a MacBook Pro carrying 64 gigabytes of memory, a machine that should not have been able to hold more than a sliver of that model at once. The weights were not shrunk to fit. Developer Marco Bambini says the full parameter set, an 896-expert version of Kimi K3, ran intact.
Do not read that as a substitute for a GPU cluster. Bambini’s own benchmark generated sixteen tokens in 49 seconds, a rate of roughly 0.32 tokens per second, well short of anything usable for a live chat interface or an agent loop today. WASTE proves the model can load and answer questions on ordinary hardware. It does not yet prove the setup is fast enough to use.
Bambini published WASTE on July 30 as the opening release in what he describes as a longer effort to run large models on hardware that people already own. Kimi K3, whose weights AI Insiders covered when the lab shipped them and its wider open stack days earlier, is the project’s first fully supported target.
The approach depends on how K3 actually computes each answer. The model routes every token through only 16 of its 896 experts per layer, so nearly all of its parameters sit unused at any given instant. WASTE keeps the shared trunk resident in memory and leaves the expert weights on the laptop’s internal drive, retrieving them only when the router calls for them.
That shifts the constraint from memory size to storage speed. Each token pulls roughly 17 gigabytes of expert data from disk, so Bambini’s team benchmarked their drives before committing to a layout. The laptop’s internal SSD delivered 12.78 gigabytes per second, while an external drive connected over USB managed only 0.94 gigabytes per second, a gap caused by the USB connection and not the drive itself. The converted model now lives on internal storage.
Memory tuning carried its own surprises. WASTE’s planner defaults to a 46 gigabyte budget out of the laptop’s 64, holding back a margin because larger allocations pushed the operating system into memory compression and paging that slowed answers rather than speeding them up. A Metal backend built for the machine’s GPU underperformed the CPU path by 22 percent in testing, so it ships disabled by default.
The size penalty is visible against WASTE’s other supported model. Kimi-Linear, a 48 billion parameter relative of K3, fits into a 19 gigabyte container and answers at about 8.9 tokens per second on the same hardware, nearly 28 times the pace WASTE reaches running K3. That gap, not the trillion-parameter headline, is the real measure of what streaming weights from disk costs today.
WASTE ships under an Apache 2.0 license, hosted on GitHub, with support already built in for macOS, Linux and Windows. Teams evaluating on-device inference for privacy or cost reasons should treat this release as evidence that trillion-parameter models can load locally, not as a benchmark to plan a product launch against until throughput closes the gap with smaller, cache-friendly models like Kimi-Linear.
Marco Bambini detailed the WASTE engine and its Kimi K3 benchmarks in a post on his Substack newsletter on July 30, 2026.
