An open-source Swift + Metal engine that runs Google's Gemma 4 26B model on any Apple Silicon Mac in about 2 GB of RAM — by streaming only the experts each token needs from the SSD instead of loading the whole model.
Last verified · 2026-07-30 · by Moe Ameen
TurboFieldfare is an open-source inference engine that runs Google's Gemma 4 26B model locally on an Apple Silicon Mac while using only about 2 GB of RAM. It is written in Swift with Metal for GPU work, released under the Apache 2.0 license, and it surfaced publicly as a Show HN post by its creator. The specific model it targets is Gemma 4 26B-A4B — the mixture-of-experts variant that has roughly 4 billion active parameters out of 26 billion total per token — quantized to 4-bit.
The reason it can fit a 26B-class model into a couple of gigabytes is the whole point, and it is worth stating accurately. A mixture-of-experts model only uses a small subset of its weights ("experts") for any given token. TurboFieldfare keeps just the shared part of the model — roughly 1.35 GB — and the key/value cache resident in RAM, and streams only the routed experts each token actually needs off the SSD on demand. It uses bounded parallel disk reads with an expert cache and overlaps GPU compute with those reads, so the SSD does the storage job that RAM normally does. In effect it trades unified memory for fast flash storage, which is exactly the resource an M-series Mac has in abundance.
The performance is honest about the tradeoff. The project reports roughly 5–6 tokens per second on an 8 GB M2 MacBook Air and roughly 31–35 tokens per second on an M5-class MacBook Pro — usable for drafting and chat, faster on newer silicon, not a replacement for a datacenter GPU. It also ships an experimental OpenAI-compatible local server, so tools that speak the Chat Completions API — including streaming and function/tool calls — can point at it as if it were a hosted endpoint, except everything runs on your laptop with no network call.
Two caveats to set expectations. It is Apple-Silicon-only and requires a recent stack — macOS 26 with Metal 4 on an arm64 Mac — so it is not a cross-platform runtime. And it is an inference engine, not a content tool: it runs the model and returns text. There is no captioning, no image or video rendering, no design layer, and no publishing. It solves "run a capable model privately on my Mac for free," which is a real and useful problem, and stops there.
TurboFieldfare's edge is that it makes capable text generation free, private, and offline: the model lives on your SSD, the drafting happens in RAM you already own, and nothing you type leaves the laptop. That is genuinely valuable for the first, sensitive half of content work — brainstorming angles from a confidential client brief, drafting hooks and scripts in bulk, roughing out a blog on a flight. What it produces, though, is plain text. It renders no image, no video, no carousel, and it publishes to nothing.
Kompozy is the half that turns that private draft into finished, distributed content. The concrete workflow: draft locally in TurboFieldfare — a week of hooks, a script per video, a blog outline — then paste those into Kompozy, which generates the media the engine can't: persona and avatar video, face-locked persona photos, multi-slide carousels via HyperFrames, quote graphics, infographics, and long-form blogs and newsletters. Kompozy rewrites the copy in your own voice through a Persona Brief, burns in branded captions, reframes every clip vertically per platform, and schedules and publishes the whole set across eight social platforms plus blog and email from one review pipeline with autopilot. The split is clean and useful: TurboFieldfare is the free, offline drafting desk that keeps your raw thinking on your own hardware; Kompozy is the production line and distribution network that ships it as 18 on-brand formats.
TurboFieldfare is an open-source (Apache 2.0) inference engine, written in Swift and Metal, that runs Google's Gemma 4 26B model locally on an Apple Silicon Mac using about 2 GB of RAM. It does that by keeping the model's shared core and KV cache in memory and streaming only the routed experts each token needs from the SSD.
Gemma 4 26B-A4B is a mixture-of-experts model — only about 4 billion of its 26 billion parameters are active per token. TurboFieldfare keeps just the shared ~1.35 GB core plus the KV cache in RAM and streams the specific experts each token routes to from the SSD on demand, overlapping those disk reads with GPU compute. It trades unified memory for fast flash storage.
The project reports roughly 5–6 tokens/second on an 8 GB M2 MacBook Air and about 31–35 tokens/second on an M5-class MacBook Pro. It requires an Apple Silicon (arm64) Mac running macOS 26 with Metal 4; it is Apple-Silicon-only and does not run on Intel Macs, Windows, or Linux.
No. It is an inference engine — it runs the Gemma 4 model and returns text. Its output is text only; it generates no images or video, has no captioning or design layer, and publishes nowhere. To turn its drafts into finished media and post them across platforms you pair it with a content engine like Kompozy.
Privacy, cost, and offline access. Everything stays on your Mac, so sensitive material never leaves the machine; there is no per-token API bill, so batch drafting is effectively free; and it works with no internet connection. The tradeoffs are speed versus a datacenter GPU, Apple-Silicon-only support, and that you handle setup yourself.