+0
Daily
991
stars
64
forks
15
watchers
1
prs
143
issues
243
releases
README
Native macOS app for the vMLX inference server on Apple Silicon
Looking for a native Swift macOS app or Swift inference engine? See osaurus.ai.
---
MLX Studio is a full-featured macOS app that wraps the vMLX inference server — large language models, vision models, and image generation served via an OpenAI-compatible HTTP API on Apple Silicon. Built on Apple's MLX framework for GPU-accelerated inference, self-hosted on the user's machine.
Screenshots
Chat Interface
Image Generation
Developer Tools
Menu Bar Tray
Download
Download the latest DMG from Releases
1. Download vMLX-X.Y.Z-tahoe-arm64.dmg for macOS Tahoe (recommended), or
vMLX-X.Y.Z-sequoia-arm64.dmg for macOS Sequoia
2. Open the DMG and drag vMLX to your Applications folder
3. Launch from Applications or Spotlight
All releases are signed and notarized by Apple for Gatekeeper compatibility.
Quick Start
1. Launch MLX Studio from Applications
2. Pick a model -- browse and download from Hugging Face directly in the app
3. Chat -- start a conversation, attach images, or use tool calling
4. That's it. No Python setup, no terminal, no configuration files.
Features
- Chat -- Multi-turn conversations with streaming, markdown rendering, and code highlighting
- Vision -- Send images to multimodal models (Qwen-VL, LLaVA, Pixtral, and more)
- Reasoning -- Native support for thinking/reasoning models (DeepSeek-R1, QwQ, GLM-Z1)
- Tool Calling -- Function calling with structured output for agent workflows
- Image Generation -- Text-to-image with MLX-powered diffusion models
- Text-to-Speech -- TTS via Kokoro with multiple voices
- Speech-to-Text -- Whisper-based transcription
- OpenAI-Compatible API -- HTTP server on localhost for any compatible client
- Menu Bar Mode -- Runs in the tray for quick access
- Developer Tools -- Model inspection, conversion, quantization, and diagnostics
- Session Management -- Save, restore, and organize conversations with SQLite persistence
- JANG Quantization -- First-class support for JANG mixed-precision quantization formats
Engine Highlights
- Continuous Batching with PagedAttention for efficient memory use
- Speculative Decoding for faster generation on supported models
- KV Cache Quantization to fit larger contexts in limited memory
- Prefix Caching for faster repeated prompts
- Hybrid Model Support (Mamba/SSM + Transformer architectures)
System Requirements
| Requirement | Minimum |
|---|---|
| macOS | 14.0 Sonoma or later |
| Chip | Apple Silicon (M1, M2, M3, M4 -- any variant) |
| RAM | 8 GB (16 GB+ recommended for larger models) |
| Disk | ~500 MB for the app; models vary (1--50 GB each) |
What's Included
MLX Studio bundles everything needed to run the inference server:
| Mode | Description |
|---|---|
| Chat | Conversational UI with vision, reasoning, and tool support |
| Server | OpenAI-compatible API server (/v1/chat/completions, /v1/responses) |
| Image | Text-to-image generation Top issues
#117[Bug] MiniMax-M2.7 JANGTQ_K: gibberish output regression introduced between v1.5.32 and v1.5.336 comments · 2026-05-16#75Add ModelScope mirror support for users in mainland China5 comments · 2026-04-15#31Feature Request : Support for HTTP/SSE Transport and "java" command in STDIO for MCP settings5 comments · 2026-03-25#121Cannot load JANGQ-AI/Nemotron-3-Super-120B-A12B-JANG_2L3 comments · 2026-06-04#132LFM2.5-VL: converted bundles rename `feed_forward.w1/w2/w3` to `gate_proj/up_proj/down_proj` (are these models in scope?)1 comments · 2026-08-13#67Future Request if not already implemented: HotSwap of LoRa Adapters1 comments · 2026-04-13#149Feature request: System One decision-model endpoint (POST /v1/systemone)0 comments · 2026-11-07#148Royal star0 comments · 2026-11-01
Latest release
vMLX 1.6.84 v1.6.84 2026-11-07vMLX 1.6.84 Changes - Publish vMLX 1.6.83 updater manifest (b20e7513) - feat(ricursive): Ripple-65B runtime and DFlash2 text lane (f6bd17ff) - fix(dflash2): text lane parity, rotating snapshot aliasing, drafter window (0b33f432) - fix(ripple): lane-owned max_tokens, history reasoning separators, launch test (abdfeec3) - perf(dflash2): lane matmul on text targets, exact Ripple verify attention (aff7cfeb) - perf(ripple): sliding-window SSD deltas; no thinking/effort controls (9b6d00ad) - test(dflash2): pin the lane matmul on text-only DFlash2 targets (3e5a51d9) - perf(dflash2): keep the finis