← back to mlx.studio
MLX Studio for macOS
Version 1.6.55 · Apple Silicon (arm64) · Tahoe native, Sequoia compatible
Code-signed and notarised — Developer ID: ShieldStack LLC
Released 2026-09-06
01
Pick your build
Two builds of the same release. If you're on macOS 26 Tahoe, or on an M5, take the
native one — it ships MLX wheels compiled for Tahoe. Otherwise take the compatibility build.
tahoe native556 MB
macOS 26 Tahoe
Native macosx_26 MLX wheels. The right choice on Tahoe and on M5.
sha2566bf9d9ce7a37226376e3d323204fef56677ced4a43182ad2d148351ea1311f19
↓ download .dmg
sonoma · sequoia534 MB
macOS 14 / 15
Use this build on Sonoma and Sequoia.
sha256f092526715880aef4a1a640d79bce96cdd30f7409c3207b7c473761f0e44059b
↓ download .dmg
Verify a download before you open it: shasum -a 256 ~/Downloads/vMLX-1.6.55-tahoe-arm64.dmg
should print the digest shown above.
────────────────────────────────────────────────────────────────────────────────────────────────────────
02
Install it
01
Open the DMGDouble-click the DMG to mount it. The app is code-signed and notarised by Apple, so there is no Gatekeeper warning.
02
Drag to ApplicationsDrag MLX Studio into the Applications folder, then close the DMG window and eject the disk image.
03
Launch MLX StudioOpen it from Applications or Spotlight. On first launch it installs the vMLX engine automatically, in one click.
04
Pick a modelSearch and download any MLX model from Hugging Face inside the app, or point it at models you already have. Optimised builds are published at huggingface.co/JANGQ-AI.
05
StartCreate a session and hit Start. Chat, use the agentic tools, or connect over the OpenAI-compatible API at 127.0.0.1:8000.
────────────────────────────────────────────────────────────────────────────────────────────────────────
03
Requirements
| Requirement | Detail |
|---|
| Platform | macOS 14.5 or later (Apple Silicon). Remote endpoints work on macOS 14+. |
| Chip | Apple Silicon — M1 or later. Intel Macs are not supported. |
| Memory (minimum) | 8 GB unified memory |
| Memory (recommended) | 16 GB or more, for 7B–20B models |
| Disk | Roughly 2 GB for the app, plus whatever your models need |
More unified memory means larger models: 16 GB comfortably handles up to about 20B
parameters, 32 GB about 35B, 64 GB about 70B, and 192 GB reaches 400B+ MoE models. The
engine's KV cache quantisation (q4/q8) pushes each of those further.
────────────────────────────────────────────────────────────────────────────────────────────────────────
04
What's in the DMG
One self-contained app. No Python, no pip, no Docker, no command line. On first launch
it installs the vMLX engine for you.
01The appStreaming chat, 20+ agentic tools, image generation and editing, voice, vision, collapsible reasoning, and a Hugging Face model browser.included
02The enginevMLX, installed automatically on first launch — five-layer caching, continuous batching, speculative decoding.included
03The API serverOpenAI- and Anthropic-compatible endpoints on 127.0.0.1:8000, plus remote endpoint support.included
────────────────────────────────────────────────────────────────────────────────────────────────────────
05
If something goes wrong
Open an issue on GitHub with your Mac model, macOS version and what you were doing —
that's usually enough to reproduce it.