Deltafin Publishes a 20x Throughput Curve Over Six Days Running Full Kimi K3 on an M1 Max Laptop

GitHub·medium signalXBlueskyLinkedInCopy link

Deltafin (gavamedia, 641 stars, Rust, created 2026-07-28, MIT) is a single native binary that runs all 16 routed experts of Kimi K3 with no pruning, using draft models for speculative decoding while K3 itself verifies every token. The README logs a dated benchmark history on one M1 Max: 0.0141 tok/s on July 27, 0.1311 on July 28, 0.2660 on July 30, and 0.2847 today, a 20x gain in six days. That published-regression-curve style is rare in inference repos and makes it the cleanest public record of how fast the local-K3 optimization race is actually moving.