macMLX is a native macOS application for running local large language models on Apple Silicon. It provides a SwiftUI GUI, a command-line tool called macmlx, and an OpenAI-compatible HTTP server on localhost:8000 — all powered by Apple's MLX framework via the mlx-swift-lm Swift package. It is open source under the Apache 2.0 license.

Does macMLX require Python?

No. macMLX is fully Swift-native. The default inference path uses the mlx-swift-lm Swift Package Manager dependency in-process. There is no Python runtime, no Electron, and no external binary required.

Which Macs can run macMLX?

macMLX requires macOS 14 (Sonoma) or later and an Apple Silicon chip — M1, M2, M3, or M4 (any variant: base, Pro, Max, Ultra). It does not run on Intel-based Macs.

Is macMLX free and open source?

Yes. macMLX is licensed under Apache 2.0 and its source code is hosted at https://github.com/magicnight/mac-mlx. The application itself is free to download and use.

How does macMLX compare to LM Studio, Ollama, and oMLX?

LM Studio uses Electron and GGUF-format models. Ollama is CLI-only and uses GGUF. oMLX runs MLX natively but exposes a web UI rather than a native Mac app. macMLX is the only tool that combines a genuinely native SwiftUI macOS app, a Swift-native CLI, and MLX-native inference in one project. All three expose an OpenAI-compatible HTTP API.

Can I use macMLX with Cursor, Continue, Cline, Raycast, Zed, or Open WebUI?

Yes. macMLX runs an always-on OpenAI-compatible server at http://localhost:8000/v1 whenever a model is loaded (or whenever `macmlx serve` is running). Any tool that accepts a custom OpenAI base URL — Cursor, Continue, Cline, Raycast, Zed, Open WebUI, and similar — works without modification. The API key can be any non-empty string.

Which LLMs can macMLX run?

macMLX runs any MLX-format model from Hugging Face. The mlx-community organisation publishes hundreds of pre-quantised models including Qwen3, Gemma 3, Llama 3, DeepSeek, Phi 3, Mistral, and SmolLM. Starting v0.4, macMLX will also support 16 vision-language model architectures (Qwen2.5-VL, Qwen3-VL, Gemma-3, SmolVLM, Paligemma, Pixtral, and more) via the MLXVLM library already in the dependency tree.

Does macMLX collect telemetry or send data to the cloud?

No. macMLX runs entirely on-device. There is no telemetry, no cloud inference, and no account system. The only outbound network activity is model downloads from Hugging Face (or a user-configured mirror such as hf-mirror.com).

Where does macMLX store its data?

All persistent data lives under ~/.mac-mlx/ in the user's real home directory: models (weights), conversations (chat history JSON), model-params (per-model overrides), downloads (resume-data for interrupted downloads), benchmarks (one JSON per run), logs (Pulse LoggerStore), settings.json, and macmlx.pid. The location is intentionally a dotfile under $HOME so that macOS App Sandbox's dotfile exemption allows read/write without user-selected file entitlements.

Why is the DMG not signed or notarized?

The project does not currently have a paid Apple Developer account (tracked in issue #19). Until that changes, the DMG is unsigned and Gatekeeper will block it on first launch. Users can bypass with `xattr -cr /Applications/macMLX.app` in Terminal, or via Finder right-click → Open.

What is on the roadmap?

v0.3.6 is a small maintenance patch (macmlx --version auto-bump, macmlx search subcommand, binary slim-down, CLI log-level flags). v0.4.0 adds Vision-Language Model support via MLXVLM. v0.5 adds LoRA adapter loading and conversation/dataset export. v0.6 adds speech I/O (WhisperKit ASR + AVSpeechSynthesizer TTS). v0.7 introduces an opt-in Community Benchmarks service aggregating submissions by chip × model × quantisation × macOS version for a public leaderboard.

Which inference engine does macMLX use?

macMLX uses Apple's mlx-swift-lm Swift Package (pinned at 3.31.x) in-process. This includes MLXLLM for text models and MLXVLM for vision-language models (used starting v0.4). Subprocess-based engines (SwiftLM for 100B+ MoE models, Python mlx-lm) were evaluated but closed as not planned because macOS App Sandbox blocks spawning external binaries.

macMLX — Native macOS LLM inference for Apple Silicon, powered by Apple MLX

Why macMLX

Not another Electron wrapper.

The only tool that gives newcomers a real SwiftUI app AND gives developers a real CLI — both talking to the same in-process MLX engine.

Feature	macMLX	LM Studio	Ollama	oMLX
Native macOS GUI	SwiftUI	Electron	—	Web UI
MLX-native inference	✓	GGUF only	GGUF only	✓
Command-line interface	✓ Swift-native	—	✓	✓
Resumable downloads + HF mirrors	✓	partial	partial	—
OpenAI-compatible API	✓ always on	✓	✓	✓
Zero Python required	✓	✓	✓	—

Three surfaces, one core

Built like a macOS app should be.

MacMLXCore is the Swift SPM package that owns all inference. The GUI, the CLI, and the HTTP server are thin shells over the same protocol.

macMLX.app

SwiftUI, macOS 14+, Apple Silicon only.

Onboarding wizard picks engine + model directory
HuggingFace browser with resumable downloads
Conversation sidebar: rename, delete, rewind-to-here
Parameters Inspector (⌘⌥I) — per-model persistence
Benchmark tab, Logs tab, Menu bar extra
Sparkle EdDSA-signed auto-update

macmlx (CLI)

swift-argument-parser · native ANSI dashboards.

pull · list · run · serve · ps · stop
Honours preferredEngine + per-model parameters from GUI
Unicode progress bars with sub-cell precision
Boxed startup banner · coloured REPL prompt
PIDFile coordination · graceful SIGTERM
JSON output on every command for scripting

OpenAI-compatible API

Hummingbird 2 · localhost:8000/v1 · SSE streaming.

POST /v1/chat/completions · GET /v1/models · GET /x/status
Cold-swap loads any local model on demand (v0.3.3)
Concurrent swap requests serialised actor-side
Drop-in for Cursor, Continue, Cline, Raycast, Zed, Open WebUI
Real RSS reported on /x/status (no fake 0 bytes)

Current release

v0.3.5 — native ANSI dashboards, cleaner deps.

Released April 17, 2026. Removed SwiftTUI + PulseUI (both Swift 6 incompatible on macOS), rebuilt the three CLI dashboards (pull, serve, run) on a tiny in-house ANSI toolkit.

v0.3.5 2026-04-17

Added

CLITerm ANSI toolkit: colour, TTY detection, U+258x sub-cell progress bars, box-drawing headers
Unicode progress bar on macmlx pull with speed + ETA
Boxed startup banner on macmlx serve; tidier REPL header on macmlx run

Removed

SwiftTUI (unmaintained, Swift 6 strict-concurrency incompatible)
_PullDashboardView / _ServeDashboardView / _ChatTUIView stub types
PulseUI + PulseProxy (ConsoleView is #if !os(macOS)-gated)

Status

CLI tests: 16/16 green
MacMLXCore tests: 90/90 green
Issue #18 closed — real live CLI dashboards landed via CLITerm

Roadmap

Shipped. Shipping. Next.

Two products, one shared MacMLXCore. Twelve releases since v0.1. Each row below links to the actual tag or plan document.

Shipped

v0.1.0

Initial MVP

Native SwiftUI GUI · menu bar · CLI (serve / pull / run / list / ps / stop) · HuggingFace downloader · OpenAI-compatible API · Sparkle auto-update · memory-aware onboarding.

v0.2.0

Download + chat polish (10 issues)

Resumable downloads survive cancel + app quit · HF mirrors · Markdown rendering · message edit/regenerate · Parameters Inspector (⌘⌥I).

v0.3.0

Benchmark feature + cross-cutting gap-fix

Local benchmark tab (prefill + generation TPS, TTFT, peak RSS, history, Share-to-Community issue template) · 4 CRITICAL + 3 HIGH + 3 MEDIUM gap-fixes from an independent code review · bilingual README.

v0.3.1

Five UX fixes

macmlx list segfault fixed · chat banner flicker fixed · Markdown paragraph breaks preserved · manually-copied models auto-appear · chat toolbar model switcher actually works · max-tokens TextField replaces click-heavy Stepper.

v0.3.2

Conversation sidebar + rewind-to-here

Collapsible sidebar lists saved conversations · inline rename · delete with confirmation · right-click any message → Rewind drops every later message.

v0.3.3

API cold-swap model loading

/v1/chat/completions now auto-loads any locally-downloaded model by ID · concurrent swaps serialised actor-side · OpenAI-style 404 model_not_found error shape.

v0.3.4

Logs tab (native over Pulse LoggerStore)

SwiftUI Table with time / level badge / category / message · search field + level picker · Clear button wipes the on-disk store.

v0.3.5

Native ANSI CLI dashboards; SwiftTUI + PulseUI removed

In-house CLITerm toolkit replaces stub-linked SwiftTUI · PulseUI dropped (ConsoleView is iOS/iPadOS-only) · Logs tab keeps working via direct LoggerStore access.

In progress

v0.3.6

Small maintenance patch (~2h)

macmlx --version auto-bumped via release.yml sed step (fixes perpetual 0.1.0 bug)
macmlx search <query> subcommand — reuses HFDownloader.search, --sort=downloads|likes|recent, --limit, --json
Binary slim-down: strip -S + dynamic Swift stdlib (~60 MB → ~45 MB)
CLI --log-level + --log-stderr flags so Pulse logging is visible from the terminal

Next minor

v0.4.0

Vision-Language Model support (#23)

MLXVLM already ships in our mlx-swift-lm dependency — 16 VLM architectures available without a new dependency: Qwen2.5-VL · Qwen3-VL · Gemma-3 · SmolVLM / SmolVLM2 · Paligemma · Pixtral · Idefics3 · FastVLM · LFM2-VL · glm_ocr · mistral3.

ChatMessage gains images: [ImageAttachment] (backward-compatible decode)
ModelFormat.mlxVLM tag · auto-detected from config.json's model_type
Image picker in ChatInputView (NSOpenPanel + drag-drop + paste)
OpenAI multimodal content-array parsing in HummingbirdServer · 10MB/image · 4 images/msg cap
Images copied to ~/.mac-mlx/conversations/<uuid>/images/ so conversations stay portable

Later

v0.5

LoRA adapters + conversation/dataset export

Drop in existing HuggingFace LoRA adapters (no training UI) · export conversations to JSONL for fine-tuning datasets.

v0.6

Speech I/O — ASR + TTS

WhisperKit (Core ML) for mic input in chat — upstream mlx-swift-lm doesn't ship audio models yet · AVSpeechSynthesizer for reading assistant replies aloud.

v0.7

Community Benchmarks service NEW

Today the Benchmark tab's Share to Community button pre-fills a GitHub issue. Tomorrow: an opt-in remote endpoint receives submissions, aggregates by chip × model × quant × macOS version, and serves a public leaderboard — the data page inside the app and on this website. Inspired by omlx.ai's community benchmarks.

Submission: POST /v1/benchmarks with BenchmarkResult JSON + anonymised HardwareInfo
Opt-in — no data leaves the Mac unless the user explicitly clicks Share
Public browsable leaderboard on this website — filter by chip family, memory, model family, quant
GitHub-issue submission continues as a fallback for users who prefer not to run the remote service

Deferred / blocked

#19

Signed + notarized DMG

Needs a paid Apple Developer account. Until then, DMG is unsigned — Gatekeeper asks users to run xattr -cr on first launch.

#12 / #13

Subprocess-based engines (SwiftLM, Python mlx-lm)

Closed as not planned — App Sandbox policy blocks spawning external binaries. Reopenable if sandbox policy is revisited or a Swift-native 100B+ MoE path appears.

#20

Homebrew tap for the CLI

Scheduled around v0.3.6–v0.4 once the CLI tarball lands as a release asset.

Benchmarks — today & tomorrow

From shared-issue to live leaderboard.

The Benchmark tab already measures prefill + generation tok/s, TTFT, peak RSS, load time, and stores history locally. Sharing is a one-click GitHub-issue pre-fill today. v0.7 plans to turn that submission into data you can query.

Today — v0.3.0

Share to Community

Result is encoded into a pre-filled GitHub issue using benchmark_submission.yml. Review before submit; nothing leaves the Mac until you click Create Issue.

benchmark Benchmark · Qwen3-8B-4bit · M3 Max 64GB

## System
Chip: Apple M3 Max
Memory: 64 GB
macOS: 15.4
Engine: MLX Swift (mlx-swift-lm 3.31.3)

## Result
Model:       Qwen3-8B-4bit
Prefill TPS: 182.4
Gen TPS:     45.1
TTFT:        0.32 s
Peak RSS:    18.2 GB
Load time:   4.8 s
Runs:        3 (median)

v0.7 — Community Benchmarks

Browsable leaderboard (preview)

Remote endpoint stores your opt-in submission, aggregates globally by chip × model × quantisation × macOS version, and publishes a filterable table on this site — plus inside the app so the Benchmark tab can show you how your Mac compares.

Chip Mem Model Gen TPS TTFT N

M4 Max 128G Qwen3-8B-4bit 62.3 0.21 s 47

M3 Max 64G Qwen3-8B-4bit 45.1 0.32 s 118

M3 Pro 36G Qwen3-8B-4bit 31.0 0.45 s 72

M2 Max 64G Qwen3-8B-4bit 28.4 0.51 s 34

M1 Max 32G Qwen3-8B-4bit 22.1 0.64 s 28

Mockup. Real data once v0.7 ships.

Architecture

One protocol. Three consumers. No leaky abstractions.

macMLX.app SwiftUI

macmlx CLI

HTTP clients Cursor · Continue · curl

↓

MacMLXCore Swift SPM · @MainActor / actors · Swift 6 strict concurrency

↓

InferenceEngine protocol

HummingbirdServer localhost:8000/v1

↓

MLXSwiftEngine

mlx-swift-lm 3.31.3 · MLXLLM + MLXVLM (v0.4) · in-process

↓

Apple Silicon Metal · ANE · Unified Memory

Data at `~/.mac-mlx/`

~/.mac-mlx/
├── models/              # weights (default; changeable)
├── conversations/       # chat history JSON
├── model-params/        # per-model overrides
├── downloads/           # resume-data for interrupted downloads
├── benchmarks/          # benchmark results (one JSON per run)
├── logs/                # Pulse LoggerStore
├── settings.json        # user preferences
└── macmlx.pid           # CLI daemon coordination

A dotfile under real $HOME — macOS App Sandbox's dotfile exemption lets a sandboxed app read/write without user-selected.read-write entitlements or security-scoped bookmarks, while staying visible to power users in Finder.

Quickstart

Running a 4-bit 8B model in 60 seconds.

1

Install

Download the DMG from Releases, drag macMLX.app to /Applications. On first launch, run xattr -cr /Applications/macMLX.app in Terminal (DMG is not notarized yet — issue #19).
2

Onboard

The setup wizard picks ~/.mac-mlx/models as the default model directory and selects the MLX Swift engine. Memory check warns if your Mac has less than the model's recommended RAM.
3

Download + chat

Open the Models tab, switch to Hugging Face, search for a model (try mlx-community/Qwen3-8B-4bit), click Download — progress bar with live speed and ETA, resumable across app quits. Load it from the Local tab, then head to Chat.

# install dev tools, clone, build the CLI
git clone https://github.com/magicnight/mac-mlx && cd mac-mlx
brew bundle
swift build --package-path macmlx-cli -c release

# download, run, serve
macmlx pull mlx-community/Qwen3-8B-4bit     # resumable
macmlx list                                   # local models
macmlx run Qwen3-8B-4bit "Hello, world"      # single prompt
macmlx run Qwen3-8B-4bit                       # interactive REPL
macmlx serve                                   # OpenAI API on :8000
macmlx ps                                      # is serve running?
macmlx stop                                    # graceful SIGTERM

# v0.3.6 preview
macmlx search qwen3 --sort likes --limit 10  # new in v0.3.6
macmlx serve --log-level debug --log-stderr  # new in v0.3.6

# anything OpenAI-compatible works. API key is ignored.
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen3-8B-4bit",
    "messages": [{"role": "user", "content": "Hi"}],
    "stream": true
  }'

# cold-swap: ask for any locally-downloaded model by ID
# server loads it on demand (v0.3.3+), concurrent swaps serialised
curl http://localhost:8000/v1/chat/completions \
  -d '{"model":"gemma-3-4b-it-qat-4bit","messages":[...]}'

# real RSS reported on status
curl http://localhost:8000/x/status | jq
# { "state": "ready", "model": "Qwen3-8B-4bit", "rss_gb": 18.2, ... }

Special thanks

Standing on shoulders.

macMLX wouldn't exist without these open-source projects. Click through and star them.

MLX Apple · ml-explore

Local LLMs on Apple Silicon, done the macOS-native way.

Not another Electron wrapper.

Built like a macOS app should be.

macMLX.app

macmlx (CLI)

OpenAI-compatible API

v0.3.5 — native ANSI dashboards, cleaner deps.

Shipped. Shipping. Next.

Initial MVP

Download + chat polish (10 issues)

Benchmark feature + cross-cutting gap-fix

Five UX fixes

Conversation sidebar + rewind-to-here

API cold-swap model loading

Logs tab (native over Pulse LoggerStore)

Native ANSI CLI dashboards; SwiftTUI + PulseUI removed

Small maintenance patch (~2h)

Vision-Language Model support (#23)

LoRA adapters + conversation/dataset export

Speech I/O — ASR + TTS

Community Benchmarks service NEW

Signed + notarized DMG

Subprocess-based engines (SwiftLM, Python mlx-lm)

Homebrew tap for the CLI

From shared-issue to live leaderboard.

Share to Community

Browsable leaderboard (preview)

One protocol. Three consumers. No leaky abstractions.

Running a 4-bit 8B model in 60 seconds.

Install

Onboard

Download + chat

Standing on shoulders.

MLX

mlx-swift-lm

swift-transformers

Hummingbird

Sparkle

Pulse

swift-argument-parser

Swama

oMLX

SwiftLM

WhisperKit

SwiftTUI

Local LLMs on Apple Silicon,
done the macOS-native way.