TurboFieldfare 2026: a 26B Gemma in 2GB RAM on any M-series Mac

TurboFieldfare is an open-source Swift and Metal inference engine for Google's Gemma 4 26B-A4B, a 26B-parameter mixture-of-experts model that activates only a few billion parameters per token. Instead of holding the full 14.3GB of 4-bit weights in memory, it keeps a 1.35GB shared core and the FP16 KV cache resident and streams just the routed experts each token needs from SSD, landing around 2GB of RAM usage.

Andrey Mikhaylov, an iOS and Metal engineer, built it as a bet that SSD bandwidth could substitute for RAM on modern Apple Silicon. He released it as a Show HN on 2026-07-29 with measured results - about 5-6 tokens/s on an 8GB M2 MacBook Air and 31-35 tokens/s on an M5 MacBook Pro - plus a streaming installer that repacks about 15GB of Hugging Face weights directly into a custom format without ever staging the full checkpoint on disk.

The post drew 919 points and 345 comments, commenters benchmarked it on M1 MacBooks and MacBook Neo hardware, and coverage followed on Creative AI News and ifun.de. Creative AI News counted 800+ HN points and 1,600+ GitHub stars within the first day; the repository showed about 410 forks by 2026-09-02.

The project is explicitly a specialist: its author positions it against Ollama, LM Studio, and llama.cpp as the runtime that pushes one model onto low-memory Macs rather than a general-purpose replacement. As of 2026-09-02 it remains free (Apache-2.0), independent, and unmonetized, with iPhone and iPad ports listed as next.