Swift engine runs Gemma 4 on iPhone

Swift-gemma4-core is a newly open-sourced Swift inference engine for running Gemma 4 natively on Apple Silicon and iOS, built after the author hit compatibility issues with existing MLX-based libraries. The project focuses on an offline, on-device experience and claims support for Gemma 4’s newer quirks, including partial rotary embeddings, cross-layer KV cache behavior, and prompt/template handling that previously broke decoding. The author says it already runs on a real iPhone with a relatively small memory footprint, but prefill latency is still high and they are asking the community to help optimize the bridge, tensor mapping, and allocations.