25 released projects · Apple silicon
Native apps, focused commands, and practical tools for running image, language, speech, music, and agent models on your Mac.
# native Qwen Image generation
$ QwenImageCLI --prompt "paper city"
Denoising 45% · ETA 1m08s
✓ saved qwen-image.png
› Runs here. Stays here.
Released for Apple silicon
There is no umbrella toolkit to install. Each card points to the project's own installer, exact Mac download, or native package command—or directly to its App Store page.
A native Qwen Image app and CLI for generation, editing, LoRAs, quantization, and layered image decomposition.
Open the DMG to install the native Mac app.
open QwenImageApp.dmgA local Qwen3.6 coding agent with one-command installation, project memory, and downloadable GGUF models.
$ curl -fsSL https://raw.githubusercontent.com/mzbac/qwen-agent/main/install.sh | shInstalls the latest agent and downloads its default Qwen model.
qw-agentA native macOS chat agent for Gemma 4 12B with multimodal input, local tools, and no server or API key.
$ curl -fsSL https://raw.githubusercontent.com/mzbac/g4-agent/main/install.sh | shInstalls the native binary and downloads the Gemma 4 model.
g4-agentA compact local DeepSeek V4 Flash agent distributed as a native Apple-silicon release with optional model setup.
$ curl -fsSL https://raw.githubusercontent.com/mzbac/ds4-agent/main/install.sh | shInstalls the release binary with optional local-model setup.
ds4-flash-agentA local MLX inference server wrapper for agentic harnesses, with single-machine and distributed serving modes.
$ uvx kooka-server serve --model mlx-community/MiniMax-M2.1-4bituvx fetches the package and starts the server in one step.
An OpenAI-compatible Swift server for text, vision-language models, embeddings, KV cache, and prompt cache.
Unzip the precompiled Apple-silicon server, then run it.
swift-mlx-server --model hf/model/id --port 8080Pipeline-parallel local LLM inference across multiple Macs, including shard servers and an OpenAI-compatible API.
$ pip install mlx-shardingInstalls the published Python package and command-line tools.
Native file and live-microphone transcription using Mistral’s Voxtral Mini Realtime model through MLX Swift.
Unzip the native CLI, make it executable, and run it.
./VoxMLXCLI --helpLow-latency local text-to-speech with offline generation, streaming playback, stdin mode, and quantization.
Unzip the precompiled Apple-silicon command and run it.
vibevoiceCLI speak "Hello from my Mac."A Swift/MLX port of NVIDIA PersonaPlex for full-duplex, live microphone-to-model-to-speaker conversation.
Unzip the native CLI, make it executable, and run it.
./personaplex --helpNative text-to-music generation with remix, repaint, extraction, composition, and audio-code workflows.
Unzip the precompiled music-generation CLI and run it.
./ACEStep15CLI --helpA prebuilt Apple-silicon OCR command for extracting local documents and images with zai-org/GLM-OCR.
Unzip the signed native OCR command and run it.
glm-ocr ocr --input scan.pngA Swift/MLX command-line port of DeepSeek OCR 2 for single images, batches, and local model directories.
Extract the Apple-silicon archive, then run the CLI.
DeepSeekOCR2CLI --image scan.pngLocal document-to-Markdown OCR with grounding prompts, image export, and a signed native command-line release.
Extract the signed native OCR command, then run it.
./DeepSeekOCRCLI ocr image.pngTurn local files and directories into a streaming MLX chat session with persistent prompt KV caches.
Unzip the ready-made command instead of building from source.
files-to-chat-swift ./notesNative Z-Image Turbo generation with LoRA, ControlNet, quantization, and a precompiled Apple-silicon binary.
Unzip the precompiled Apple-silicon image CLI and run it.
ZImageCLI -p "A mountain at dusk" -o output.pngText-to-image and image-to-image FLUX.2 inference with model caching, quantization, and prompt upsampling.
Unzip the native FLUX.2 CLI, make it executable, and run it.
./flux2-cli --helpA native FLUX.1 command for text-to-image and Kontext image editing across Schnell, Dev, and Kontext models.
Unzip the native FLUX command and run it from Terminal.
./flux --prompt "A quiet coastal road"The underlying MLX Swift implementation of FLUX.1, supporting pre-quantized weights for native integrations.
$ .package(url: "https://github.com/mzbac/flux.swift.git", from: "0.1.3")Add this dependency entry to your Swift package manifest.
swift package resolveA local FLUX image-generation API with text and image workflows, real-time progress, authentication, and storage.
Open the DMG to install the packaged local server.
./ImageMatesServerA Swift package for local BERT, Qwen, and multilingual embedding models with batching and concurrent inference.
$ .package(url: "https://github.com/mzbac/mlx.embeddings.git", from: "0.1.0")Add this dependency entry to your Swift package manifest.
import MLXEmbeddingsA private macOS menu-bar app for capturing and analyzing screenshots entirely on-device with local vision models.
Unzip ScreenMate.app and move it to Applications.
open ScreenMate.appA VS Code agentic coding assistant with a headless runtime, workspace tools, subagents, and OpenAI-compatible local servers.
Download the extension package, then install it in VS Code.
code --install-extension lingyun-2.5.3.vsixA private native Mac AI art studio for local text-to-image, image-to-image, LoRA workflows, editing, and gallery management.
Free Mac App Store app · runs locally on Apple silicon.
An on-device iPhone document assistant for private PDF chat, OCR, multi-document search, and citation-backed answers.
Free iPhone app · document processing stays on-device.
From release to run
No generic toolkit is required. Use the project's own installer, download its Mac build, copy its package command, or open the App Store.
Install
install.sh · pip · uvx · SwiftPMOr download
*.macos.arm64.zip · *.dmg · *.vsixThen run
local process · local files · local outputWhy local
Prompts, files, and outputs remain on your Mac. Airplane mode is a feature, not a failure state.
The stack is tuned around unified memory and Apple's GPU architecture—not merely made compatible with it.
Switch models, inspect performance, and integrate locally without accounts, quotas, or surprise usage bills.
Featured release
The native app is the quickest way into the collection. The same repository also ships a full CLI for generation, editing, layered output, quantization, and profiling.
NATIVE MAC APP
QwenImageApp.dmgmacOS 14+ · Apple siliconPR #2: progress UI + layered speed ↗