A free, open-source macOS app that runs Alibaba's Qwen3-ASR speech model entirely on Apple Silicon — private file transcription plus system-wide dictation, with no cloud, account, or API key.
Last verified · 2026-07-29 · by Moe Ameen
Qwen Scribe is a free, open-source macOS app that does two things privately and entirely on your own machine: it transcribes audio and video files, and it acts as a system-wide dictation tool. It's an independent community project by a developer who goes by VladUZH, published on GitHub under the Apache-2.0 license, and it surfaced publicly as a Show HN in 2026. Its own one-line pitch is "private, local transcription and system-wide dictation for Apple Silicon."
The engine underneath is Alibaba's Qwen3-ASR speech-recognition model, run through the mlx-qwen3-asr port so it executes on the Mac's Metal GPU rather than a cloud API. You choose between two model sizes: a 1.7B model tuned for accuracy that needs roughly 3.4 GB of unified memory, or a 0.6B model tuned for speed that needs around 1.2 GB. The first transcription downloads the chosen weights once; after that the app works offline. Nothing — audio or transcript — leaves the Mac, and there's no account, API key, or subscription anywhere in the flow. The project is explicit that it is not affiliated with, endorsed by, or sponsored by Alibaba Cloud, the Qwen team, or Apple.
The two workflows are distinct. For files, you drag audio or video into a local web interface and get back an editable transcript with word-level timestamps, automatic language detection (with an option to force a language and add vocabulary hints), and an SRT subtitle export. For dictation, you hold the right Command key while focused in any text field, speak, and release — Qwen Scribe transcribes on device and inserts the text at your cursor, with a small heads-up display for listening/transcribing states that's designed not to steal focus. A local, searchable transcript history lets you reopen, delete individually, or clear everything.
Two honest caveats. It's Apple Silicon only and needs macOS 14 or newer, Python 3.12+, and FFmpeg for non-WAV media — so Windows, Linux, and Intel-Mac users are out. And it's early: at the time of writing it's version 0.1.0-beta.1, distributed as a source build (you compile it locally with `make app`; signed, notarized binaries are slated for a later release), and saved transcripts are stored as readable, unencrypted JSON.
The most underrated thing Qwen Scribe unlocks isn't transcription — it's dictation as a capture method. Because you can hold right Command and talk into any text field, you can narrate a content idea the second it lands: the hook for a video, the angle for a carousel, the rough outline of a newsletter, all spoken aloud and typed on device without a keyboard or a cloud round-trip. The problem is that a spoken draft is still just raw words in a text box. Kompozy is what turns that captured voice into finished, on-brand, published content — so the pairing is "talk it, then ship it."
Concretely: dictate your idea or a loose script anywhere on your Mac with Qwen Scribe, then paste that text into Kompozy as a source. From that one spoken draft Kompozy generates across formats — a Persona or HeyGen avatar video that delivers your script on camera, Clipped Shorts, brand-exact Carousels via HyperFrames, Photo Posts, Quote Graphics, a Blog Article, an Email Newsletter, and Text Posts — every piece governed by your Persona Brief so a rushed, spoken-out-loud thought comes out sounding like your brand. Kompozy then reframes each piece per destination and schedules and publishes across eight social platforms plus blog and email through Autopilot and a per-post review queue. Qwen Scribe removes the keyboard from ideation; Kompozy removes everything between the idea and a week of published posts.
Qwen Scribe is a free, open-source (Apache-2.0) macOS app for Apple Silicon that transcribes audio and video files and provides system-wide voice dictation. It runs Alibaba's Qwen3-ASR model locally on the Mac's Metal GPU via mlx-qwen3-asr, so audio and transcripts stay on your machine — no cloud, account, or API key. It's an independent project by VladUZH, not an official Alibaba or Apple product.
Yes, after a one-time setup. The first transcription downloads your chosen model weights (a 1.7B accuracy model needing ~3.4 GB of memory, or a 0.6B speed model needing ~1.2 GB); once the dependencies and weights are cached, it runs fully offline with nothing sent to any server.
An Apple Silicon Mac running macOS 14 or newer, Python 3.12 or newer, and FFmpeg for non-WAV audio or video. Dictation also needs microphone, input-monitoring, and accessibility permissions. There's no Windows, Linux, or Intel-Mac support, and at v0.1.0-beta.1 you build it from source with make app rather than installing a signed binary.
Focus any text field, hold the right Command key, speak, then release. Qwen Scribe transcribes on device and inserts the text at your cursor, showing a small non-intrusive heads-up display for listening and transcribing. A local searchable history keeps your past transcripts, which you can reopen or delete.
No. It produces a transcript or dictated text and stops there — it doesn't design posts, generate video, caption for a feed, or publish. To turn a dictated draft into finished content, pair it with a content engine like Kompozy, which generates video, carousels, images, a blog, and a newsletter in your brand voice and publishes them across platforms.