MLX Chat — Intelligence in a box

WHAT'S INSIDE

Download the app. Pick a model. Chat. No inference server to stand up, no client to point at it — the ceremony is gone; the machinery stays.

01 — CATALOG

A curated model catalog

Qwen, Gemma, Llama, Phi, DeepSeek — quantized for Apple Silicon. Or import any MLX-compatible Hugging Face repo by id.

02 — SYSTEM MODEL

Apple Intelligence, built in

Zero downloads, zero setup. And when Apple ships its next, more powerful on-device model, MLX Chat inherits it — no code change.

03 — REASONING

Thinking, disclosed

Reasoning models stream their think phase into a collapsible disclosure — with a per-model toggle and a live tok/s readout.

04 — ARTIFACTS

Artifacts, rendered live

The model writes HTML or SVG; it renders beside the chat in a sandboxed web view. Markdown, syntax highlighting, and LaTeX included.

05 — STORAGE

An honest model manager

Download with progress and cancel. Delete cleanly. Disk usage that adds up. Everything lives in ~/Models, in the open.

06 — MEMORY

Auto-Unload After

Idle models give their RAM back — after 2 minutes, 10 minutes, an hour, or never. The Metal cache is cleared on the way out, so it actually comes back.

TWO ENGINES · ONE PROTOCOL

One model in RAM.
Behind one protocol.

ENGINE 01 — MLX

Local weights, unified memory.

Quantized models load straight into Apple Silicon's unified memory and stay hot between turns.

KV-cached ChatSession reused across turns — no re-prefilling the conversation
History trimmed to 70% of the context window before prefill
Metal buffer cache capped at 20 MB — no gigabytes parked after unload

MLX Chat conversation — thinking disclosure, LaTeX math, a Swift code block, and a live tok/s readout

Manage Models — downloaded Qwen models with size and context badges, plus curated model families

ENGINE 02 — APPLE INTELLIGENCE

The OS is the model manager.

The FoundationModels framework hands MLX Chat the system's on-device model. Nothing to ship, nothing to download, nothing to babysit.

Transcript-seeded LanguageModelSession, fresh per turn
Context-window and guardrail errors mapped to human messages
Asks the OS for its context size — future Apple models are a free upgrade

BOTH ENGINES — ONE MENU

Switch engines mid-thought.

Apple Intelligence and your local models share one picker. Choose a model and it starts warming immediately — by the time you've typed the next message, the weights are hot.

One click from Apple Intelligence to any downloaded model
Switch-warming loads ahead of your next send, confirmed with a quiet "Switched to …" toast
Download Models… lives right in the menu — no settings detour

MLX Chat model picker — Apple Intelligence and downloaded Qwen models in one dropdown, with Download Models entry

GET STARTED

Up and running
in two minutes.

Download the DMG or build from source. On Apple Intelligence–enabled Macs, chat works the moment the app opens — no model download needed.

macOS App Recommended

Drag to Applications. Chat with Apple Intelligence out of the box, or pull a local model from the built-in catalog.

Download DMG

From source

Requires an Apple Silicon Mac, macOS 26 with Xcode 26, and XcodeGen.

# clone and generate
git clone https://github.com/abayomipopoola/mlx-chat
cd mlx-chat && xcodegen generate

# build (mlx-swift ships a plugin + macros)
xcodebuild -scheme MLXChat -configuration Release \
  -skipPackagePluginValidation \
  -skipMacroValidation build