Three Projects Get Kimi K3 Running on Laptops

Three independent projects put the full 2.78-trillion-parameter Kimi K3 model on single machines within a week of Moonshot's July 26 weight release.

Moonshot's own guidance calls for 64 or more accelerators to run K3. That's data-center hardware, not a desk.

Getting the same weights onto a 64GB MacBook Pro or a single M1 Max changes who can open the model up at all. Daily use is still out of reach for any of them.

WASTE, a GitHub project with 1,366 stars, keeps a 27.28GB trunk resident in memory. It streams the rest of K3's experts from NVMe using 3-bit residual vector quantization. That shrinks the model's footprint from a published 1.42TB down to 982GB on disk. It runs at 0.45 to 0.62 tokens per second.

Its own README logs a strange result. Growing the expert cache from 17.32GB to 29.32GB raised the cache hit rate from 36.2% to 41.3%. Throughput fell eightfold anyway, because a cache hit turned into a page fault.

Deltafin, a Rust project with 641 stars, published the clearest paper trail. It logged 0.0141 tokens per second on July 27, climbing to 0.2847 by August 3, a 20x jump in six days on one M1 Max.

kimi-k3-in-c goes the other direction: 176KB of plain C99, no BLAS, no framework, no GPU. It runs at 8.24GB peak RSS and 32.69 seconds per token.

These speeds aren't usable for anything real. They're proof the model runs at all, not a product.

WASTE's own numbers make that case better than any argument could. A bigger cache made the model slower, not faster, because it stopped hitting RAM and started hitting the page fault handler. Watch whether the next multi-trillion-parameter model gets a laptop port this fast, or whether K3's tricks turn out to be a one-off.