TurboFieldfare

TurboFieldfare is a custom Swift and Metal runtime for running the instruction-tuned Gemma 4 26B-A4B model on Apple Silicon Macs, including MacBooks with 8 GB of RAM. It streams the model experts needed for each token from SSD instead of loading the full model into memory.

The project includes a Swift command-line interface, a native Mac app, a streaming installer, and a sibling decode-service executable. Users clone the repository, build it with Swift, download and repack the pinned model, then load the model and generate text locally. The model download is about 15 GB, with an optional image pack of about 1.1 GB.

TurboFieldfare targets macOS 26 with Metal 4 and Swift 6.2. The repository reports 5.1–6.3 tokens per second on an 8 GB M2 MacBook Air and 31–35 tokens per second on a 24 GB M5 Pro.

Highlights

• Runs Gemma 4 26B-A4B with about 2 GB of weights and a 4K KV cache
• Uses Swift and Metal rather than MLX or llama.cpp
• Supports Apple Silicon Macs with 8 GB of RAM
• Includes a native Mac app, CLI, installer, and decode service
• Available as an open GitHub repository