jjang-ai/mlxstudio

What it solves

MLX Studio is a native macOS desktop application that allows users to run large language models (LLMs), vision-language models (VLMs), and image generation models locally on Apple Silicon. It eliminates the need for cloud services, API keys, and data transmission, providing a private, all-in-one environment for local AI.

How it works

Built on the vMLX Engine and Apple's MLX framework, the app leverages Metal for high-performance inference. It provides a user-friendly interface for downloading models from HuggingFace, managing sessions, and interacting with AI through a chat interface or an OpenAI-compatible API server. It also includes advanced quantization techniques, such as JANG adaptive mixed-precision, to optimize model size and memory usage on Mac hardware.

Who it’s for

It is designed for macOS users with Apple Silicon (M-series chips) who want to run a wide variety of AI models locally for chatting, coding, image generation, or as a local API backend for other applications.

Highlights

• Extensive Model Support: Supports over 65 model families, including text LLMs, vision models, and multimodal omni models.
• OpenAI & Anthropic Compatibility: Provides a drop-in replacement API server for OpenAI and Anthropic SDKs.
• Agentic Workflows: Includes 26+ built-in tools for file I/O, shell execution, web search, and MCP (Model Context Protocol) support.
• Local Image Generation: Integrated support for Flux and Z-Image models for local image creation and editing.
• L5 Cache Stack: A five-layer caching system (including prefix and paged KV caches) to maximize throughput and memory efficiency.
• Advanced Quantization: Features JANG adaptive mixed-precision and "Smelt Mode" for partial expert loading of MoE models to save RAM.
• Integrated Model Tools: In-app tools for converting GGUF models to MLX format and performing various quantization tasks.