Get started

The Mac app

A native SwiftUI app around the same engine. Chat, a model manager that knows what fits, one-click agent setup, and a lab that shows speculative decoding working on your own prompts.

The Race above ran plain decoding and DFlash 2 on Qwen3.8-27B (8-bit) side by side, then compared the token ids: identical output, 4.58× faster on that prompt.

Install#

With Homebrew:

brew tap ARahim3/mlx-dspark https://github.com/ARahim3/mlx-dspark
brew trust arahim3/mlx-dspark          # Homebrew 6+: third-party taps need explicit trust
brew install --cask mlx-dspark
xattr -dr com.apple.quarantine /Applications/mlx-dspark.app

The app isn't notarized yet. The last line clears the quarantine flag once. Alternatively, open it, then use System Settings › Privacy & Security › Open Anyway after the first launch. On Homebrew 5 or older, brew install --cask --no-quarantine mlx-dspark replaces the trust and xattr steps.

Or download the DMG from Releases (the app-v* tags) and drag it to Applications.

First launch#

The app sets up its own private engine runtime the first time, which takes about 2–4 minutes once. It doesn't touch Homebrew's Python or any virtual environment of yours, and it never uses a pip-installed engine. After that it:

keeps the engine on the latest release automatically, so engine fixes reach you without an app update;
tells you when a newer version of the app itself exists (brew upgrade --cask mlx-dspark).

What's inside#

Chat: Saved sessions, markdown and syntax highlighting, collapsible reasoning, and per-turn timing: time to first token, decode speed, tokens accepted per round, and how much of the prompt the cache served.
Models: The measured pairs with their speedups, anything already on disk (including LM Studio's MLX downloads and your own model folders), and any Hugging Face repo. It answers "will this fit my Mac?" before you download, and swaps models without restarting the server.
Agents: One-click setup for Claude Code, Codex, OpenCode, pi and any OpenAI-compatible app: the exact config for each, and a round-trip test.
Lab: Race runs several decoders on the same prompt and checks the lossless claim token by token. Live streams per-round acceptance. Curves shows this Mac's measured cost curves. This Mac shows your measured memory bandwidth, the plain-decoding ceiling for the loaded model, how far above it speculation runs, and macOS memory pressure and swap.
Settings: Decoding controls, a fixed engine port, serving on your local network with an API key, and your model folders.
Menu bar: A live gauge with tokens per second, model memory, and acceptance.

Serve other devices from the app#

Settings › Local server has the same two switches as mlx-dspark serve --host 0.0.0.0 --api-key <key>: listen on every interface, and require a key. With a key set, every route except /health needs Authorization: Bearer <key> or x-api-key.

Build it yourself#

The app is a SwiftPM package with no Xcode project, and builds with the Command Line Tools alone. See apps/MacApp/README.md.