jjang-ai/mlxstudio

What it solves

MLX Studio provides a user-friendly, all-in-one macOS application for running large language models (LLMs), vision models, and image generation locally on Apple Silicon. It removes the technical barriers of setting up Python environments, terminal commands, or configuration files, allowing users to deploy AI models directly from Hugging Face on their own hardware.

How it works

The app wraps the vMLX inference server and leverages Apple's MLX framework to enable GPU-accelerated inference. It bundles its own Python 3.12 environment and provides an OpenAI-compatible HTTP API on localhost. For performance, it uses advanced techniques like continuous batching with PagedAttention, speculative decoding, and KV cache quantization to optimize memory and speed.

Who it’s for

It is designed for macOS users with Apple Silicon (M1-M4) who want a plug-and-play experience for running multimodal AI models locally without needing to manage complex software dependencies.

Highlights

• Multimodal Capabilities: Supports chat with vision (multimodal models), reasoning models, and text-to-image diffusion models.
• Integrated Tooling: Includes built-in text-to-speech (Kokoro), speech-to-text (Whisper), and tool calling for agent workflows.
• Developer Suite: Features tools for model inspection, conversion, quantization (including JANG mixed-precision), and diagnostics.
• Ease of Use: Offers direct Hugging Face model downloads, a menu bar tray mode, and session management via SQLite persistence.