TurboFieldfare

Running a 26-billion-parameter mixture-of-experts model like Gemma 4 usually assumes a machine with tens of gigabytes of free memory, which rules out most consumer Macs. TurboFieldfare, a native Swift and Metal project from developer Andrey Mikhaylov (drumih), attacks that assumption directly. Instead of loading the full 14.3 GB checkpoint into RAM, it keeps only a 1.35 GB shared core plus the FP16 KV cache resident, and streams the routed mixture-of-experts weights from SSD on demand. The project's own framing is blunt: memory got expensive, so it gave a 26B model a roughly 2 GB budget. For anyone who has watched a large model simply fail to load on an 8 GB laptop, that pitch is immediately legible.

The mechanics are specific to Apple's platform. TurboFieldfare uses MLX-style affine 4-bit quantization together with custom Metal 4 kernels, and prefill runs in 128-token chunks to keep throughput reasonable while streaming. Because only the routed experts for the current token are pulled from disk, and reads happen through bounded parallel I/O rather than one weight at a time, the SSD becomes an extension of RAM rather than a hard bottleneck. The installer itself streams model weights directly from Hugging Face without ever materializing a full local checkpoint, which keeps disk usage close to what is actually needed rather than requiring the entire model to be downloaded first.

The natural audience is Apple Silicon owners, particularly on 8 GB or 16 GB Macs, who want a genuinely large model like Gemma 4 26B running locally without buying more hardware. It ships as a CLI, an OpenAI-compatible server, and a native Mac app, so it fits both scripted workflows and people who want a point-and-click local chat client. The project is arm64-only and requires macOS 26, so it is squarely a tool for people already on recent Apple hardware rather than a cross-platform runtime.

The caveats sit mostly in scope and evidence, not in code quality. This is Apple Silicon and macOS 26 only, with no path to Linux, Windows, or Intel Macs. The project is pre-1.0 (currently v0.9.0), moving through releases roughly every few days, which signals active development but also a still-settling API. The streamed model weights themselves are governed by Google's own Gemma license, separate from the project's Apache-2.0 code license, so redistribution terms differ from the runtime itself. And the headline throughput numbers, roughly 5 to 6 tokens per second on an 8 GB M2 MacBook Air and 31 to 35 tokens per second on a 24 GB M5 Pro, are the project's own measurements rather than independently reproduced benchmarks.

The verdict: TurboFieldfare is a credible, actively maintained answer to a real problem, running a 26B model within roughly 2 GB of memory on Apple Silicon, backed by a real Metal-kernel implementation rather than a thin wrapper. Adopt it if you own an 8 to 24 GB Apple Silicon Mac and want Gemma 4 running locally today, accepting that you are on a pre-1.0 release. It is too early if you need cross-platform support, a frozen API, or performance numbers verified outside the project's own testing.