TurboFieldfare runs Google's instruction-tuned Gemma 4 26B-A4B in about 2 GB of RAM on any Apple Silicon Mac, including 8 GB machines, by never loading the full model at all. The repository has collected 6,162 stars and 371 forks since its 17 July 2026 creation and was last pushed on 18 August 2026. It is an independent research project by Andrey Mikhaylov, an iOS and Metal engineer, and the README states plainly that it is not affiliated with, sponsored by, or endorsed by Google — the model is Google's, the runtime is not.

The mechanism is expert streaming rather than compression alone. The installed model occupies about 14.3 GB on disk, of which TurboFieldfare keeps only the shared 1.35 GB core plus an FP16 KV cache resident, then streams from SSD only the experts a given token actually routes to. Gemma 4 26B-A4B has 26B total parameters with roughly 3.88B active per token, which is what makes the trade viable. Weights are MLX affine 4-bit at group 64, with an 8-bit router and 4-bit shared and routed experts.

Per layer, Metal computes attention and the router from resident weights; the CPU takes the router's top-8 expert IDs and plans against that layer's 16-slot LFU cache, filling misses with bounded parallel pread calls straight into Metal-visible buffers. While those reads are in flight Metal computes the resident shared-expert branch, then combines the shared and routed outputs. Prompt prefill runs in chunks of up to 128 tokens so one fetched expert can serve multiple rows. KV storage is FP16, using bounded circular storage for 25 sliding-window layers and linear storage for 5 full-attention layers.

The published numbers are measured rather than projected, and the README frames them as reference points rather than ceilings: 5.1–6.3 tok/s decode on an 8 GB M2 MacBook Air and 31–35 tok/s on a 24 GB M5 Pro, with a community benchmark guide for contributing results from other machines. Behind that sits an unusually disciplined evidence trail — a curated experiment record of 103 measured results across kernels, caching, I/O, prefill and decode, including the plausible ideas that failed and the early results that reversed under stronger validation.

The Swift package exposes six products: the runtime library, a native SwiftUI/AppKit Mac app, a one-shot decode service, a CLI, a loopback OpenAI-compatible Chat Completions server, and a streaming installer. The installer is itself part of the memory story — it never materializes the full source checkpoint, instead streaming byte ranges from a pinned Hugging Face revision and repacking them directly into the .gturbo layout as they arrive, accepting the result only after manifest and file hashes validate.

The scope limits are strict and clearly stated. It is model-specific rather than a wrapper around MLX or llama.cpp, so it runs the pinned Gemma 4 26B-A4B checkpoint and nothing else. It requires macOS 26 with Metal 4, Swift 6.2 and Xcode 26, is arm64-only, and does not support older macOS or Metal versions. It is text-only — no images, audio or video — and the app and CLI neither expose nor execute tools, though the loopback server will return model-produced tool calls for a client to authorize. That server has no remote authentication or TLS and is meant to stay on loopback. Licensed Apache-2.0, with model weights downloaded separately and governed by their own terms.