Deltafin
A single native binary that runs the full, unpruned 2.8-trillion-parameter Kimi K3 model on consumer hardware, using small draft models only to speed up generation, never to reduce quality.
Deltafin runs the complete Kimi K3 model, all 16 routed experts and every token, on consumer hardware rather than the roughly 16-node, multi-terabyte-VRAM infrastructure Moonshot recommends for it. Its stated rule is that speed must never come from reducing model quality: small draft models are allowed to guess ahead, which is where much of its speed comes from, but K3 itself checks every guess and nothing reaches the user without K3's own approval.
The project frames itself as research into how far consumer hardware can be pushed to run a model of this size, rather than a finished product, and publishes ongoing throughput benchmarks on an M1 Max laptop as it improves. It targets machines on the order of a high-end workstation rather than typical consumer laptops.