CoreML-LLM
A Swift package for running LLMs such as Gemma 4, Qwen3.5 and Liquid AI's LFM2.5 on iPhone using CoreML, targeting Apple's Neural Engine so the GPU stays free for the rest of the app.
CoreML-LLM runs on-device language models on the Apple Neural Engine rather than the GPU, aimed at battery-friendly, server-free inference on iOS 18+ and macOS 15+. It supports models including Gemma 4 (E2B/E4B, text plus image, video and audio), Qwen3.5, Qwen3-VL, FunctionGemma, EmbeddingGemma and Liquid AI's LFM2.5, adding the package and calling CoreMLLLM.load(repo:) with a registered model id or a full Hugging Face path.
A model downloads on first use and its on-device bundle is reused on later calls, with streaming output, multi-turn chat, and image, video and audio input documented in the package's Quick Start guide. The project positions itself against MLX Swift, recommended instead when maximum GPU throughput matters more than keeping the GPU free. Swift, MIT licensed.