macMLX
A native macOS app, CLI and always-on API for running local LLMs on Apple Silicon through MLX, with the inference engine itself implemented in Swift rather than wrapping a Python core.
macMLX runs local large language models on Apple Silicon using Apple's MLX framework, with a SwiftUI app for newcomers, a command-line tool for developers, and an always-on OpenAI-compatible API for everything else, all with no cloud calls and no telemetry. Unlike some comparable native macOS apps, its inference engine runs in-process in Swift rather than fronting a Python core, and the whole application is roughly 50 MB with no Python required.
Its engine supports text models and 16 vision-language-model architectures, including dedicated OCR-model recognition, up to roughly 70B parameters, with continuous batching that raises aggregate throughput 2.5 to 3.2 times under concurrent clients, a tiered RAM-and-SSD prompt cache with longest-common-prefix reuse across agent turns, and speculative decoding.
It installs by downloading a DMG from GitHub Releases; the DMG is not yet notarized, so Gatekeeper's quarantine attribute needs clearing on first launch. It requires macOS 14 or later on Apple Silicon (M1 to M4).