zw5/ane-fieldfare (ANE Fieldfare) is a local inference engine that runs Gemma 4 26B-A4B on Apple Silicon. It is written in Swift and released under the Apache-2.0 license. There is no formal release yet, and it targets people who want to run the model directly on their own devices.
This repository splits the accelerators in two. Input processing, that is prefill, is handled by the Apple Neural Engine (ANE), a block dedicated to neural network operations, while model computation and speculative decoding are handled by the GPU. Speculative decoding means sampling several candidates for the next token in advance and verifying them in one pass. It reads memory-mapped model files and uses the operating system's file cache as a store for expert weights. The repository includes an installer and server, Metal kernels, and tests for model handling and communication.
That said, this division of labor may not be stable. The README notes that the Neural Engine path depends on a non-public Apple framework, so a macOS update could break it. The performance figures are also hard to verify independently from the README's measurement description alone. With one star and a creation date of October 20, 2026, this repository is closer to a candidate worth trying than a settled tool.