Skip to main content
models

What the First Local Model Download Does on Your Mac

Vemoir Team
Vemoir Team Mac meeting records by Aurum Avis Labs · Oberägeri, Switzerland
7 min read

Hero photo by sasun1990 on Unsplash

A local model download is the install you start on purpose: the speech model that will transcribe, or the language model that will write notes, documents, and answers. The app shows download size, storage, what the model is for, and whether it fits this Mac. Progress can resume after a failure or a quit. Vemoir does not silently download, select, or substitute a model while you are recording or while a meeting is processing.

This is not a guide to working with the network off. That is a different question. Here the subject is the download itself: what arrives, how you see it finish, and what does not get replaced behind your back.

What the first local model download is

Setup introduces local processing, then asks only for the permissions the capture setup needs. It then asks you to choose transcription and a local language model. Those are different files. Parakeet v3 transcribes 25 European languages. Whisper Turbo, Medium, and Large-v3 are the other speech choices, each with their own coverage. None of those write the minutes. Qwen3 4B is the language model aimed at most Macs. gpt-oss-20b is the larger one for high-memory Macs. They write notes, documents, and Ask. They do not transcribe audio.

You can skip a cloud key in onboarding. A key is not required for the download. Internet is required to fetch the files. After the models you chose are installed, a later meeting can run without the network. That later step is why the download matters. The download screen is where you see it happen.

The model library is where you return: download, inspect, select, remove. Readiness, storage, hardware compatibility, and lifecycle stay visible. A meeting run uses the exact speech model and language model selected for that run. It does not pick a smaller model mid-meeting because the disk was tight.

  1. Check free space

    Plan for roughly 10 GB or more depending on models.

  2. Download once on Wi-Fi

    Models cache on disk for offline runs.

  3. Pick Parakeet or Whisper

    Match language coverage before a long trip.

Resume, failure, and quitting

If the download stops, you get a failure and a way to recover. Progress survives quitting and relaunching the app. Denied permissions, a network drop, and a failed download are states you can leave and come back to. You can reopen setup from Settings, and you can reset it.

What you should not see is a completed meeting that used a model you never chose. If a file is missing, processing does not quietly fetch a different one and continue as if nothing changed. Fix the download, then run the meeting again.

Integrity checks and license notices ship with the local model files. You do not need to paste a URL. You do need free space. Vemoir needs 10 GB of free storage as a baseline, and long meetings are more comfortable with 16 GB of memory. The model screen is still the place that shows the size of the file you are about to fetch. Do not treat a blog post as the storage gauge.

Speech model and writing model are both downloads

You downloadIt doesIt does not
Parakeet v3 or a Whisper variantTranscribe audio on the MacWrite the minutes by itself
Qwen3 4B or gpt-oss-20bWrite notes, documents, and Ask on the MacTranscribe the recording
Nothing yetSetup waitsInvent a model and start recording on it

Apple Intelligence, where the Mac has it, can help with cleanup and artifacts without being the speech model. It is not a substitute download for Parakeet or Whisper. If you also add an OpenAI or Anthropic key, that key is text only, stored in the Keychain, and off until you activate it. Saving a key does not download a speech model, and it does not become a silent backup when a local file fails.

Where skipping a download is better

The local model download is the better choice when this Mac will transcribe and write without sending audio away. It is the worse choice when:

  • You only have one short file and you do not want to install anything. The transcribe recording tool runs in the tab, up to 15 minutes, three successful runs a week, with no speaker labels. The model for that tool downloads into the browser, which is a different install from the Mac app’s library.
  • You are on an Intel Mac, Windows, or a phone. The app is Apple silicon, macOS Sonoma 14.2 or later. No download path changes that.
  • You wanted the larger writing model and this Mac is not high-memory. Choose Qwen3 4B. Do not start gpt-oss-20b and hope the app swaps it for a lighter one during the meeting. It will not.
  • Disk space is already gone. Free space first. A resumed download still needs room to finish.

Speaker separation does not have a separate consumer download you shop for in this screen in place of the speech model. It runs on the Mac as part of processing. Unknown speakers stay anonymous until you opt in to a profile.

The controls are in capture.

Apple Intelligence and the downloaded writing models

Parakeet or Whisper and Qwen3 or gpt-oss-20b are files you install on purpose. Apple Intelligence, where the Mac has it, can help with cleanup without replacing a speech download. What it may touch in Vemoir is in Apple Intelligence for meeting notes.

Once both models show ready, you can record offline. For the next live meeting, start Vemoir from the menu bar.

After a failed download on a train

You started Parakeet v3 on hotel Wi-Fi and closed the lid. Reopen Vemoir on a network you trust: progress should remain, not a surprise Whisper file marked ready. Resume or use recovery. Do not record “just this once” hoping a cloud transcript will fill the gap. There is no such fallback. When the speech model shows ready, pick the writing model (usually Qwen3 4B, or gpt-oss-20b on a high-memory Mac). That pair is what the meeting uses, not whatever finished last in the background.

Frequently asked questions

What is the first local model download?

The speech model and the language model you choose in setup, with size and device fit shown before they install. One transcribes. The other writes.

Can the download resume?

Yes. Progress, failure, and recovery are visible. Setup progress survives quit and relaunch.

Will a meeting swap models if the download is incomplete?

No. Vemoir does not silently download, select, or substitute a model during recording or post-processing.

Do I need the network forever?

You need it for setup and for model downloads. After the models you chose are installed, the meeting workflow can run offline. That offline use is a consequence of a finished download, not a different product.

Does the download include a cloud key?

No. OpenAI or Anthropic is optional, text only, and a separate activation. The model files for speech and for local writing are the download.

Can I remove a model later?

Yes. The model library lets you inspect, select, and remove compatible local models. A meeting keeps the selections captured for that run.

How is this different from the browser tool’s first visit?

The browser tool fetches a public Whisper model into the tab for a file you already have, with a 15-minute cap. The Mac app’s library is the set of models for live capture, speaker separation, and on-device writing.

What if I deny microphone access during setup?

You can recover from Settings. A denied permission is not replaced by a silent upload. Screen Recording is the permission for the call track, used for audio only, and it does not capture the screen.

Start the download you mean to keep, watch it finish, and record after the selected models show ready. That status is in capture.

models download mac
Vemoir Team

Written by

Vemoir Team

Mac meeting records by Aurum Avis Labs · Oberägeri, Switzerland

We build Vemoir, a Mac app for local-first meeting records: capture from the menu bar, transcribe on your Mac, and turn calls into minutes and action items.

Customize your cookie preferences. Essential cookies cannot be disabled as they are required for the website to function properly.

Essential Cookies

Required for basic website functionality, security, and the free in-browser tools limit (timestamps of successful runs in the current week, no content).

Always active

Analytics Cookies

Help us understand how visitors interact with our website to improve user experience. Includes Google Analytics and Microsoft Clarity session recordings.

Marketing Cookies

Used to track visitors across websites to display relevant and engaging advertisements.