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

An open-source (Apache-2.0) Swift and Metal inference engine, CLI, and native Mac app that run Google's instruction-tuned Gemma 4 26B-A4B in about 2GB of RAM on any Apple Silicon Mac, including 8GB ones.

An MoE model's routed experts could stream from SSD instead of RAM, letting a 26B-parameter model run in about 2GB on an 8GB Mac rather than a 32GB workstation.

Andrey Mikhaylov, an iOS and Metal engineer who loves on-device AI, wanted to run a model whose 14.3GB of 4-bit weights do not fit in the memory of an 8GB Mac. He posted TurboFieldfare as a Show HN on 2026-07-29, claiming about 2GB RAM usage at 5-6 tokens/s on an 8GB M2 MacBook Air and 31-35 tokens/s on an M5 MacBook Pro.

The thread took off - 919 points and 345 comments - and the maintainer answered questions in real time, sharing measured cache-hit rates and taking community results from M1 MacBooks and MacBook Neo hardware. Coverage followed on Creative AI News and German site ifun.de; the repository records 103 audited experiments behind the design, plus an optional vision companion pack for M2+ Macs. By 2026-09-02 the repository showed about 410 forks.

As of 2026-09-02 TurboFieldfare remains an independent research project: Apache-2.0 code with no company, funding, or monetization, and future work pointed at iPhone/iPad ports and more low-memory Macs.

Memory, not compute, was the wall that kept 26B-class models off consumer laptops.

Gemma 4 26B-A4B is a mixture-of-experts model, so only a small slice of parameters fires per token and the full weight set need not be resident.

Custom Metal kernels and a model-specific runtime beat wrapping general-purpose MLX or llama.cpp on this one workload.

Fast Apple SSDs made disk bandwidth a practical substitute for RAM capacity.

HN's developer audience could clone, build, and benchmark the code themselves, turning a launch claim into a community-verified result.

Writing a model-specific engine for one hard constraint can beat general-purpose runtimes, and shipping real measured code lets a single engineer borrow a whole community as a benchmark team.

As of 2026-09-02 TurboFieldfare is still in active open-source development: 103 audited experiments, a community benchmark guide, a loopback OpenAI-compatible server, and an optional vision pack, with about 410 forks reflecting an active community. The author describes it as an independent research project with no affiliation to Google, and lists iPhone and iPad ports as main future work. German site ifun.de framed it as proof that a model this size can run on 8GB Macs while noting the trade-offs: output on base hardware is deliberate and SSD speed matters.