mlx.studio download
free · macOS · apple silicon

Everything you'd use AI for.
None of it leaves the Mac.

Chat with any open model, write code with 20 agentic tools, generate and edit images with Flux, convert models between formats, and serve an OpenAI- or Anthropic-compatible API to your other apps. One app, running entirely on Apple Silicon. No API keys, no subscription, nothing uploaded.

↓ download for mac view the engine

Free forever · macOS 14.5+ · M1 through M5 · 8 GB minimum, 16 GB+ recommended

MLX Studio — Console
The console mid-reply: the reasoning block expanded, and a per-turn footer showing tokens, throughput, prompt rate, cache state and time to first token.
agentic tools
20+
File I/O, search, shell, git, web — no setup
architectures
50+
Including Mamba and SSM hybrids
image models
9
6 generation, 3 editing, all local
cost
$0
No subscription, no API keys, no account
01

Everything you need. Nothing in the cloud.

Simple enough to open and start chatting, deep enough to run a full inference stack underneath — KV cache quantisation, prefix caching, speculative decoding and 14 tool parsers, all exposed if you want them and invisible if you don't.

01 Streaming chatMulti-turn conversations with inline tool-call pills, collapsible reasoning blocks, image previews and live status. any model
02 Agentic coding20 built-in tools with native MCP support. Models read, write, search and execute directly on your machine. 20 tools
03 Image generation and editingFlux Schnell, Z-Image Turbo, Flux Dev, FLUX.2 Klein and Qwen Image for generation; Qwen Image Edit, Flux Kontext and Flux Fill for editing. 9 models
04 Vision and multimodalDrag an image into the chat. Vision models analyse it locally, with click-to-zoom previews. qwen vl
05 Reasoning blocksCollapsible chain-of-thought for DeepSeek R1, Qwen 3 and GLM. See the working, or fold it away. 4 parsers
06 VoiceText-to-speech on every response using native Mac speech synthesis. Hands-free output. built in
07 Anthropic and OpenAI APIsA native Anthropic Messages endpoint alongside OpenAI Chat and Responses. Claude Code, OpenClaw and the Anthropic SDK all connect unchanged — and you can point at remote endpoints too. both
08 Model converterGGUF to MLX inside the app, plus JANG mixed-precision quantisation profiles. No waiting for someone else to publish a conversion. gguf → mlx
────────────────────────────────────────────────────────────────────────────────────────────────────────
02

20 built-in tools. Zero configuration.

Native MCP tool calling, on by default. The model can read your files, search your code, run a command and check git state — locally, with no plugin to install and no key to paste.

File I/O

  • read_file
  • write_file
  • edit_file
  • list_dir
  • copy
  • move
  • delete

Code search

  • grep
  • glob

Shell

  • execute_command

Web

  • duckduckgo_search
  • brave_search
  • fetch_url

Git

  • git_status
  • git_diff
  • git_log
  • git_show

Utilities

  • clipboard_read
  • clipboard_write
  • current_datetime
────────────────────────────────────────────────────────────────────────────────────────────────────────
03

Generate and edit images on your own machine

Nine models, downloaded on demand and run on Apple Silicon. Give it a prompt, or give it a photo and a prompt to inpaint, transform or restyle. No cloud API, no credits.

Flux Schnell
Fastest generation · 4 steps · ~6-24 GB
Z-Image Turbo
Fast turbo generation · 4 steps · ~6-24 GB
Flux Dev
High quality · 20 steps · ~6-24 GB
FLUX.2 Klein 4B
Fast and small · 20 steps · ~4-8 GB
FLUX.2 Klein 9B
Medium quality · 20 steps · ~8-18 GB
Qwen Image
Strong prompt understanding · 20 steps · ~20-40 GB
Qwen Image Edit
Instruction-based editing · 28 steps · ~54 GB
Flux Kontext
Subject-consistent editing · 24 steps · ~6-24 GB
Flux Fill
Inpainting with a mask · 20 steps · ~24 GB
────────────────────────────────────────────────────────────────────────────────────────────────────────
04

Underneath it, the vMLX engine

MLX Studio is the app — the chat window, the tools, the model browser. vMLX is the inference engine that powers it, and it's open source. The relationship is roughly LM Studio to llama.cpp, except both halves are ours.

01 Prefix cacheRepeated parts of a conversation are computed once and reused, so a long thread doesn't get slower. layer 1
02 Paged KV cacheEvery open chat stays resident — switching between them doesn't evict anything. layer 2
03 Cache quantisationq4/q8 cuts live cache memory 4–8×, which is what makes very long contexts fit. Stored prefixes stay full precision. layer 3
04 Continuous batchingUp to 256 concurrent sequences on one loaded model, so several agents can share it. layer 4
05 Persistent disk cachePrompt computation survives an app restart — you come back to a warm start, not a cold one. layer 5
Time to first token — M3 Ultra 256 GB · Llama 3.2 3B 4-bit · against LM Studio
ContextMLX StudioLM StudioDifference
2.5K tokensShort prompt0.05 s0.49 s9.7×
10K tokensWorking context0.08 s6.12 s76×
100K tokensCold prompt processing0.65 s131 s224×

Measured on an M3 Ultra with 256 GB, Llama 3.2 3B at 4-bit, against LM Studio. The first two rows are time to first token; the last is cold prompt processing. Decode throughput is a separate measurement that depends on both context length and output predictability, so it isn't reduced to a single number here.

────────────────────────────────────────────────────────────────────────────────────────────────────────
05

What it looks like

Screens from the app running on Apple Silicon.

Chat & Images
Chat & Images — the generation and editing model pickers, each with its step count and memory footprint.
Models — find & download
Models — searching Hugging Face and pulling MLX models in the background.
Servers & API — gateway
Servers & API — one gateway in front of every loaded model, speaking OpenAI, Anthropic and Ollama.
────────────────────────────────────────────────────────────────────────────────────────────────────────
06

Get it

Download the app, or install the engine on its own if you only want the server.

the apprecommended
# download, drag to Applications, open
# chat, tools, images, converter and
# the API server, all in one
the enginepip
$ pip install vmlx
$ vmlx serve mlx-community/Qwen3-8B-4bit
# OpenAI + Anthropic compatible on
# http://127.0.0.1:8000
────────────────────────────────────────────────────────────────────────────────────────────────────────
07

Questions

What is MLX Studio?
A free macOS app for AI chat and agentic coding, running models locally on Apple Silicon. It includes 20 built-in tools for file editing, code search, shell execution and web search, image generation and editing, a model converter, and an API server — all powered by the vMLX engine underneath.
How is MLX Studio different from the vMLX engine?
MLX Studio is the app — the chat UI, agentic tools, model browser and settings. vMLX is the inference backend that powers it: caching, batching, model loading and the API layer. Roughly the relationship LM Studio has to llama.cpp, except both halves are the same project. The engine is open source and installable on its own with pip install vmlx.
Does it need an internet connection?
Only to download models the first time. All inference runs locally with no cloud connection, no API keys and no data leaving your Mac.
What Mac do I need?
Any Apple Silicon Mac — M1 or later — on macOS 14.5 or newer. 8 GB of unified memory is the minimum; 16 GB or more is comfortable. Connecting to remote endpoints works on macOS 14+.
Can I use cloud APIs as well?
Yes. Connect to OpenAI, Anthropic, Groq or any OpenAI-compatible endpoint. The agentic tools work the same whether the model is local or remote.
What does it cost?
Nothing. Free forever, no subscription, no account, no paid tier.