Qwen Scribe review (2026): honest verdict on the free, local Qwen3-ASR transcription and dictation app for Apple Silicon — accuracy, privacy, setup, and limits.
Qwen Scribe is a clean, genuinely private way to run Alibaba's Qwen3-ASR locally on an Apple Silicon Mac — free, open source, and offline, doing both file transcription (with SRT export and word timestamps) and system-wide dictation. For private speech-to-text it's an easy recommendation, with a well-chosen feature set. The honest limits: it's Apple Silicon only, it's an early beta you build from source, and saved transcripts are unencrypted JSON. If you want private transcription or dictation, it's a strong yes. If you wanted the tool that makes and publishes content from what you record, that's a different category.
Qwen Scribe is a free, open-source macOS app from an independent developer (VladUZH), shared publicly as a Show HN in 2026 with the pitch "private, local transcription and system-wide dictation for Apple Silicon." It does two things entirely on your own machine: it transcribes audio and video files, and it acts as a system-wide dictation tool you trigger by holding the right Command key. What makes it worth a review is the engine choice under the hood — it runs Alibaba's Qwen3-ASR speech model locally through the mlx-qwen3-asr port, on the Mac's Metal GPU, so no audio ever touches a cloud service.
There's no account, API key, or subscription anywhere in the flow. The first transcription downloads your chosen model — a 1.7B model tuned for accuracy (~3.4 GB of unified memory) or a 0.6B model tuned for speed (~1.2 GB) — and after that it works offline. The project is explicit that it isn't affiliated with Alibaba Cloud, the Qwen team, or Apple; it just packages a strong open speech model into a tidy local Mac app under the Apache-2.0 license.
This review scores Qwen Scribe on what it actually is: a focused, private transcription and dictation utility. It doesn't get penalized for lacking content generation or publishing, because it never claimed those — but platform support, maturity, and setup friction are scored honestly, since they're real constraints for real users. Ratings and details below reflect the app as of 2026-07-29, at version 0.1.0-beta.1.
Qwen Scribe is a macOS app for Apple Silicon that pairs local file transcription with system-wide dictation. For files, you drag audio or video into a local web interface and get an editable transcript with word-level timestamps, automatic language detection (with an option to force a language and supply vocabulary hints), and an SRT subtitle export. For dictation, you hold the right Command key while focused in any text field, speak, and release — it transcribes on device and inserts the text at your cursor, showing a small heads-up display for listening and transcribing that's designed not to steal focus. A local, searchable transcript history lets you reopen, delete individually, or clear everything, and there's a localhost-only API with Host and browser-origin checks. Under the hood it runs Qwen3-ASR via mlx-qwen3-asr on the Mac's Metal GPU, with a choice of the 1.7B accuracy model or the 0.6B speed model. It's Apache-2.0 licensed and requires macOS 14 or newer, Python 3.12 or newer, and FFmpeg for non-WAV media; dictation additionally needs microphone, input-monitoring, and accessibility permissions. At v0.1.0-beta.1 it ships as a source build — you compile it locally with make app, with signed and notarized binaries planned for a later release. Note that "Qwen" refers to the model it runs, not to any official Alibaba product.
Qwen Scribe fits Apple Silicon Mac users who want private, offline transcription and dictation and don't want to pay for it or send audio to a cloud. That's podcasters and interviewers transcribing recordings, writers and developers who think faster than they type, and anyone privacy-conscious who wants speech-to-text that never leaves the machine. Because it exports SRT with timestamps, it's also a fair fit for anyone subtitling their own footage locally. It's a poor fit for people on Windows, Linux, or Intel Macs (it simply won't run), for anyone who needs a polished, signed installer today rather than a source build, and for creators hoping a transcription app will also produce or publish content — that's a job for a content engine, not an ASR tool.
| Dimension | Score | Why |
|---|---|---|
| Transcription accuracy | 4.2 / 5 | Qwen3-ASR is a strong open speech model, and the accuracy-tuned 1.7B option is solid for clear audio; there's no independent benchmark for this specific app yet. |
| Privacy | 5.0 / 5 | Fully offline and on-device once the model is cached — no account, API key, or network call, and audio never leaves the Mac. |
| Speed & performance | 4.0 / 5 | Runs on the Metal GPU with a 0.6B speed model for fast jobs and a 1.7B model for accuracy; real speed depends on your Mac and the model chosen. |
| Dictation experience | 4.2 / 5 | Hold right Command, speak, release, and text lands at the cursor with a non-intrusive HUD — a clean, focused dictation flow. |
| Feature depth | 3.9 / 5 | Well-chosen for the category: word timestamps, SRT export, language detection, vocabulary hints, and a searchable local history. |
| Setup & install friction | 3.0 / 5 | A beta source build requiring Python 3.12+, FFmpeg, and make app — fine for developers, a barrier for everyone else until signed binaries ship. |
| Platform support | 2.5 / 5 | Apple Silicon Macs only, macOS 14+; Windows, Linux, and Intel-Mac users are shut out entirely. |
| Maturity & stability | 3.0 / 5 | Version 0.1.0-beta.1 — core transcription, history, and dictation work, but it's young and things will change. |
| Value | 5.0 / 5 | Free and Apache-2.0 licensed with no subscription; for private local ASR, the price is unbeatable. |
There's little to analyze on price because Qwen Scribe is free and Apache-2.0 licensed. There's no tier, trial, or upsell — you build it from source and use it. Your only real cost is owning an Apple Silicon Mac with enough unified memory for your chosen model (roughly 3.4 GB for the 1.7B, 1.2 GB for the 0.6B, on top of normal overhead). That makes it dramatically cheaper than paid dictation and transcription tools, and cheaper in ongoing terms than any cloud ASR service that meters minutes.
The trade you make for free is scope and setup. Qwen Scribe gives you exactly two capabilities — private local transcription and dictation — and none of the extras paid tools justify subscriptions with (team sync, hosted storage, AI rewriting, or content workflows). It also asks more of you up front than a one-click installer: at this beta stage you need Python, FFmpeg, and a build step. For a developer-leaning Mac user who wants private ASR, that's a great deal and the absence of a price tag is a feature.
For someone who expected transcription to be a stepping stone to producing content, the honest read is that no price would make Qwen Scribe the wrong tool right — it's simply not built for that, and the thing that is (an engine that generates and publishes) is a separate purchase. Judge it as free, private ASR, where it's excellent, not as a content tool, where it isn't trying to compete.
| Use case | Fit | Why |
|---|---|---|
| Private, offline transcription of recordings | Strong | On-device Qwen3-ASR means podcasts, interviews, and meetings never leave your Mac. |
| System-wide dictation on a Mac | Strong | Hold right Command in any field, speak, and text lands at the cursor — a clean, free capture method. |
| Subtitling your own footage locally | OK | SRT export with word timestamps covers straightforward subtitling without uploading anything. |
| Everyday users who want a signed one-click app | Weak | At v0.1.0-beta.1 it's a source build needing Python and FFmpeg; signed binaries aren't out yet. |
| Windows, Linux, or Intel-Mac users | Weak | Qwen Scribe requires an Apple Silicon Mac on macOS 14+, so it simply won't run for these users. |
| Storing sensitive transcripts long-term | Weak | Saved transcripts are unencrypted JSON, so secure retention needs your own disk-level protection. |
| Turning what you record into published content | Weak | Qwen Scribe ends at a transcript; making and scheduling video, carousels, and posts is a content engine's job. |
Being straight about it: Kompozy is not a Qwen Scribe alternative, and if you came here to pick a private transcription or dictation tool, Qwen Scribe is a better answer than Kompozy for that specific job. Kompozy doesn't do local speech-to-text at all. Where it fits is the step Qwen Scribe can't reach — after the transcript, and specifically for people repurposing recordings they already have.
Here's the concrete hand-off that makes the two complementary. Say you recorded a podcast episode, a webinar, or a long interview. Qwen Scribe transcribes it privately on your Mac and can hand you an SRT with word-level timestamps. Drop that recording (or its transcript) into Kompozy as a source, and it does the part Qwen Scribe doesn't: it cuts the long recording into Clipped Shorts with branded, feed-styled captions burned in, and generates a Carousel via HyperFrames, Photo Posts, Quote Graphics of the best lines, a Blog Article, an Email Newsletter, and Text Posts from the same episode — every piece held to your voice by a Persona Brief, then reframed per platform and scheduled and published across eight social platforms plus blog and email. So the honest recommendation is to use both for what each is good at: Qwen Scribe to get an accurate, private transcript of your recording for free, and Kompozy to turn that one recording into a week of published, on-brand content — rather than expecting either one to do the other's job.
For Apple Silicon Mac users who want private, offline transcription or dictation, yes — it's free, open source, and runs Alibaba's Qwen3-ASR locally on the Metal GPU, with SRT export, timestamps, and system-wide dictation. The reasons it wouldn't be worth it are simple: you're not on an Apple Silicon Mac, you need a signed one-click installer today rather than a beta source build, or you actually wanted a tool that makes and publishes content, which transcription isn't.
It runs Alibaba's Qwen3-ASR speech-recognition model locally through the mlx-qwen3-asr port on the Mac's Metal GPU. You choose a 1.7B model tuned for accuracy (~3.4 GB of unified memory) or a 0.6B model tuned for speed (~1.2 GB). The first transcription downloads the weights once, after which it works offline.
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 version, and at v0.1.0-beta.1 you build it from source with make app.
Yes. It runs entirely on-device, so audio and transcripts stay on your Mac with no account, API key, or cloud call, and it uses a localhost-only API with Host and browser-origin checks. The one caveat is that saved transcripts are stored as readable, unencrypted JSON, so secure deletion requires your own disk-level protection.
No. It's an independent, open-source community project by a developer known as VladUZH, and it states plainly that it isn't affiliated with, endorsed by, or sponsored by Alibaba Cloud, the Qwen team, or Apple. "Qwen" refers to the underlying Qwen3-ASR model it runs, not to an official Alibaba product.
Both are private, on-device Mac tools, but they differ in engine and scope. Qwen Scribe runs Alibaba's Qwen3-ASR (downloading model weights once) and does both file transcription with SRT export and dictation, on Apple Silicon. Yap downloads no model at all — it uses Apple's built-in macOS 26 speech engine — and focuses on dictation. Qwen Scribe offers more transcription features; Yap is lighter to install.
No. Qwen Scribe transcribes audio and dictates speech into text; it doesn't generate video, images, carousels, or posts, and it can't schedule or publish. If that's what you need, pair it with a content engine like Kompozy, which turns a transcript or recording into finished multi-format content and publishes it across platforms.