Open-Source Engine Runs Gemma 4 26B in 2 GB on an M2 Air

TurboFieldfare, a Swift and Metal engine released under Apache 2.0, ran to the top of Hacker News with 730 points by streaming mixture-of-experts weights from SSD instead of holding them in RAM. It keeps the ~1.35 GB shared core and KV cache resident and fetches only the experts each token routes to, giving 5.1–6.3 tok/s on an 8 GB M2 MacBook Air. The design works because ~67% of expert lookups hit a 16-slot cache. Explicit pread() beat mmap() by 8× end to end — the fault path cannot be pipelined against GPU work.