Qwen Scribe: Local Dictation on Mac with Qwen3-ASR (MLX)

Qwen Scribe runs Qwen3-ASR on Apple Silicon through MLX to deliver file transcription and system-wide dictation entirely on-device, under Apache-2.0. Memory needs are modest: roughly 3.4GB of unified memory for the 1.7B model and 1.2GB for 0.6B, plus app overhead. No cloud, no API key — hold the right ⌘ key, speak, release, and the text lands at your cursor. It hit Show HN on July 29, 2026 and passed 100 GitHub stars within two weeks of first commit.

Two distinct capabilities: push-to-talk dictation and drag-and-drop transcription of audio or video files. The first works through an OS-level global hotkey, so the same gesture works in mail, an editor, or a chat window.

One design detail deserves credit: the API is localhost-only with Host and browser-origin checks. Local AI tooling that quietly binds to 0.0.0.0 is common enough that this is a meaningful signal if you intend to run it on a work machine.

3.4GB is light next to typical local LLM footprints — comfortable to keep resident on a 16GB Mac alongside real work, which is exactly what a dictation tool needs. For the broader model-size-to-memory mapping, see local AI hardware requirements.

At v0.2.0-beta.1 only source releases exist; signed and notarized binaries are planned once Developer ID credentials are in place. Clone the repo, then build the app bundle.

make app
open "dist/Qwen Scribe.app"

To run it as a server instead, use ./run.sh and open http://127.0.0.1:8990. Python dependencies and model weights are fetched and cached on first launch, after which it runs fully offline. You can move the built app to /Applications.

Three environment variables control behavior — worth deciding up front if data location matters to your policy.

Desktop dictation needs three grants under System Settings → Privacy & Security. Knowing why each is required makes it far easier to justify internally.

- Microphone: records while the push-to-talk key is held
- Input Monitoring: detects the modifier key while another app is frontmost
- Accessibility: pastes the finished text into the focused field

Input Monitoring and Accessibility are, by nature, powerful — they let software act across other apps. The on-device design and readable Apache-2.0 source are what justify granting them. Conversely, if you evaluate a closed-source tool that also ships audio to the cloud, the same permissions deserve a much harder look.

Local transcription on Mac has been Whisper territory (MacWhisper, Handy, and others), so the real comparison axes are noise robustness, punctuation, OS coverage, and maturity — not raw accuracy claims.

One HN commenter reported roughly 2x realtime on an M1. If that is not fast enough, drop to 0.6B; if proper nouns matter more than speed, pair 1.7B with vocabulary hints. For the cloud-side dictation experience see Aqua Voice, and for wiring a recorder into an AI workflow see Plaud MCP & CLI.

Worth checking before you commit. Upstream Qwen3-ASR does support Japanese — its primary languages are listed as Chinese, English, Cantonese, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish, with language identification across 52 languages and dialects. Qwen Scribe's author, however, listed English, German, French, Russian, Italian, and Spanish as the languages verified in the Show HN thread, so Japanese quality is outside what the author tested.

Public Qwen3-ASR implementations also note extra tokenizer dependencies such as nagisa for Japanese. If you intend to use it daily in any language outside the author's tested set, run a one-minute sample of your own voice and your own terminology first and check punctuation and proper nouns before building it into a workflow.

- Saved transcripts are unencrypted, plain JSON on disk. For sensitive conversations, decide on QWEN_SCRIBE_DATA_DIR placement, FileVault, and a habit of deleting old history
- v0.2.0-beta.1, unsigned and un-notarized — you will be working around Gatekeeper. Personal evaluation before fleet rollout is the realistic path
- Native Python 3.12+ is required. If a Rosetta or older Python takes precedence in your PATH it will not run; confirm python3 -c "import platform; print(platform.machine())" returns arm64
- Accessibility and Input Monitoring grants are policy-relevant at some companies. Check internal rules before installing
- "It stays local" is the actual feature for regulated work — the HN discussion centered on clinicians who cannot send patient audio to cloud scribes. For that class of use, the architecture is the requirement

The README also states plainly that this is an independent community project, unaffiliated with Alibaba Cloud, the Qwen team, or Apple. Do not plan around vendor support.

Qwen Scribe satisfies one clear requirement — dictation that never leaves the machine — in exchange for one clear constraint: Apple Silicon only. The 3.4GB (1.7B) / 1.2GB (0.6B) footprint suits an always-resident tool, and the localhost-only API, explicit note about plaintext history, and Apache-2.0 source all help justify granting it Accessibility-level permissions on a work machine. It is still v0.2.0-beta and unsigned, and languages beyond the author's tested set are unverified, so the right next step is a sample of your own audio rather than a fleet rollout.