System requirements
About 3.7 GB of models, once. Then almost nothing.
Apple silicon Mac · macOS 14.4 or later · free. Everything below runs on your Mac; there is no server side to size.
What you need
| Mac | Any Apple silicon Mac: M1 or later, including MacBook Air. |
|---|---|
| macOS | 14.4 Sonoma or later. On macOS 26 with Apple Intelligence on, Apple’s on-device model adds a further polishing step to dictation and notes; everything works without it. |
| Memory | 8 GB works: the app downloads the same notes model plus a 1.3 GB quick model for rewrites and replies, and pauses dictation for the minute the notes are written. 16 GB or more runs the full set with room to spare. |
| Disk | About 3.7 GB for the models (about 5 GB on 8 GB Macs), made up as below. The app itself is about 65 MB. |
| Languages | English on a model tuned for it (NVIDIA Parakeet); 98 other languages on whisper, detected as you speak or pinned in Settings. |
| Network | Only for two things: downloading the models once from Hugging Face, and a daily check for updates through this site and GitHub. Nothing you say is ever sent. |
Disk, exactly
| The app | About 65 MB installed, about 25 MB to download. |
|---|---|
| English dictation model | NVIDIA Parakeet TDT 0.6B v2 · 460 MB · runs on the Neural Engine. |
| Speech model, 99 languages | whisper.cpp small · 490 MB · also handles very short replies. |
| Notes model | Qwen3.5-4B · 2.7 GB · only if you turn on meeting notes · writes the overview, key points and next steps after each meeting. |
| Quick model | Qwen3.5-2B · 1.3 GB · 8 GB Macs only · handles quick work while you wait, so the notes model isn't needed for it. |
| Speaker models | pyannote speaker separation · about 25 MB · with meeting notes · tells the people in a meeting apart by voice. |
| Your notes and transcripts | Text in an encrypted local database. A year of daily use is under 50 MB. |
| Meeting audio | Kept for 7 days by default, so a meeting can be transcribed again. It's stored uncompressed, so it's large: roughly 1 GB per hour of meeting (your microphone and the call audio). Settings › Meetings can delete it after transcribing, or keep it until the meeting is deleted. |
Memory while it works
| In the menu bar | Under 100 MB, with the dictation model kept warm so the first word is quick. |
|---|---|
| Dictating | Around 1 GB for the moment it takes. |
| Writing meeting notes | Around 3 GB for about a minute after a meeting ends. |
Battery and heat
Dictation is a fraction of a second of work per sentence and you will not notice it. An hour-long meeting is real work for the machine: expect a fanless MacBook Air to run warm, not hot, and to use battery at about the rate of a video call.
Permissions it asks for
| Microphone | To hear you. Required. |
|---|---|
| Accessibility | To type where your cursor is in other apps, and for a single-key shortcut such as Fn or Control on its own. Without it, dictation is copied to the clipboard instead. |
| Screen Recording | The permission macOS uses for computer audio, so the other side of a call can be heard. No pixels are captured. |
| Calendar, Reminders, Contacts | Optional. Names meetings from your events; adds the action items you choose to Reminders; spells the names of people you know. |
For comparison
The cloud dictation apps keep a local cache of one to two gigabytes on your Mac and still send your audio away. Mythral keeps its models on your Mac and sends nothing.