TurboFieldfare runs Gemma 4 26B on an 8 GB MacBook

Apple sells laptops with 8 GB of unified memory and a chip fast enough to run models that do not fit in it. TurboFieldfare, written by one developer, is a Swift and Metal runtime that gets Gemma 4 26B-A4B, a 14.3 GB model, running in roughly 2 GB of RAM on any M-series Mac.

What it is

A model-specific inference runtime written from scratch rather than wrapped around MLX or llama.cpp. It includes the runtime itself, a streaming installer, a CLI and a native Mac app, all in Swift and Metal. Apache-2.0. The README states the motivation in its opening line: memory got expensive, so a 26-billion-parameter model was given a 2 GB budget.

Why it showed up now

Version 0.4.3 two days before the snapshot, on a repository one month old, pushed daily. The repository also keeps a curated experiment record: 103 measured results across kernels, caching, I/O, prefill and decode, kept in the repository as a document rather than a blog post.

How it actually works

Gemma 4 26B-A4B is a mixture of experts, and the runtime exploits that shape directly. The shared 1.35 GB core stays resident along with an FP16 KV cache, and for each token only the experts that token actually needs are streamed from SSD. Nothing else is loaded, which is how the model fits into a machine whose entire memory is smaller than the checkpoint.

Being model-specific is a deliberate trade-off. A general runtime has to handle any architecture, which means it cannot assume where the expert boundaries are or how to lay them out on disk for the access pattern that follows. Giving that up buys layout and scheduling decisions tuned to one model, and it rules out every other model.

Writing it in Swift and Metal rather than adopting MLX puts the author on the hook for the kernels, which is presumably why the experiment inventory exists at all.

Try it

git clone https://github.com/drumih/turbo-fieldfare.git
cd turbo-fieldfare
swift build -c release
.build/release/TurboFieldfareMac

On first launch, choose Download and let it fetch and repack the pinned model, about 15 GB. Then Load Model and type a prompt.

Where it is weak

The name of the project should be read as its scope: this runs Gemma 4 26B-A4B. It does not run your fine-tune or the next model Google ships, and when that model is superseded, the runtime needs work to follow it.

Streaming experts from SSD on every token is sustained read traffic on the machine's internal drive, and on a Mac that drive is not replaceable. The 8 GB configuration is also the one where the streaming never stops, so it is the configuration with the least headroom for anything else you are doing.

Version 0.4.3, 52 open issues, and a single maintainer. The benchmark discipline is good, but those numbers come from the author's own hardware and methodology. The README invites contributed results because one machine is not a sample.