Installing models on a restricted network

What is actually going on

Tape ships at about 15 MB. The models are far bigger, so they are fetched on first use rather than bundled: speech models from huggingface.co, language models from the same place. Once the files are on disk, transcription and summaries never leave your Mac.

If your company runs a proxy or filter (Zscaler, Netskope, a plain firewall rule), the download can fail in one of two very different ways. Knowing which one you have decides what to do next.

Find out which one you have

Open Terminal and paste these two lines:

curl -sS -o /dev/null -w 'catalog: %{http_code}\n' \
  https://huggingface.co/api/models/argmaxinc/whisperkit-coreml/revision/main

curl -sS -o /dev/null -w 'file:    %{http_code}\n' \
  https://huggingface.co/argmaxinc/whisperkit-coreml/resolve/main/openai_whisper-tiny/config.json

Anything that is not a number, such as a hang or a certificate error, counts as blocked.

Option 1: ask IT for a read-only allowlist

This is the cleanest fix and usually an easy request, because nothing here uploads anything. Tape reads public model files, with no account and no credentials. The hosts to allow:

huggingface.co             model listings and file downloads
cdn-lfs.huggingface.co     the large weight files
cdn-lfs-us-1.hf.co         same, alternate CDN
cas-bridge.xethub.hf.co    newer storage backend for some repos

Useful wording for the ticket: "macOS app, downloads public read-only AI model files over HTTPS on first launch, no login, no data leaves the machine. Please allow the hosts above for my account." A temporary window of an hour is enough. Once the models are downloaded, Tape never needs those hosts again.

Option 2: download once from another network

The models only need to arrive once. Tether to your phone or take the Mac home, open Tape Settings → Models, and download the models you want. Back on the corporate network everything keeps working from disk.

Download at least one speech model and one language model:

• Parakeet (~600 MB) for English meetings, or a Whisper size for other languages
• Gemma 4 E2B (~2 GB) for summaries and Ask AI

Option 3: install the files by hand

This is what to do when the Mac itself can never reach Hugging Face. Get the files on a machine that can (a home laptop, a colleague's Mac that already runs Tape), then move them across by AirDrop, USB drive, or your company's own file share.

Speech: Parakeet (English, ~600 MB)

On the Mac that has it working:

cd ~/Library/Application\ Support
tar -czf ~/Desktop/parakeet.tgz FluidAudio

On the restricted Mac, quit Tape, then:

mkdir -p ~/Library/Application\ Support
tar -xzf ~/Desktop/parakeet.tgz -C ~/Library/Application\ Support

The result should look like this:

~/Library/Application Support/FluidAudio/Models/parakeet-tdt-0.6b-v2/
├── Preprocessor.mlmodelc      · folders, not files
├── Encoder.mlmodelc
├── Decoder.mlmodelc
├── JointDecision.mlmodelc
├── config.json
└── parakeet_vocab.json

Source repository, if you are fetching the files yourself: FluidInference/parakeet-tdt-0.6b-v2-coreml.

Speech: Whisper (multilingual, 75 MB to 3 GB)

Whisper needs two folders: the Core ML model itself, and a small tokenizer folder of about 3 MB that lives beside it. Copy both, or the model loads and then stops looking for words. Swap small for the size you want (tiny, base, small, medium, large-v3).

cd ~/Documents/huggingface/models
tar -czf ~/Desktop/whisper-small.tgz \
  argmaxinc/whisperkit-coreml/openai_whisper-small \
  openai/whisper-small

On the restricted Mac:

mkdir -p ~/Documents/huggingface/models
tar -xzf ~/Desktop/whisper-small.tgz -C ~/Documents/huggingface/models

~/Documents/huggingface/models/argmaxinc/whisperkit-coreml/openai_whisper-small/
├── MelSpectrogram.mlmodelc    · folders, not files
├── AudioEncoder.mlmodelc
├── TextDecoder.mlmodelc
├── config.json
└── generation_config.json

~/Documents/huggingface/models/openai/whisper-small/
├── config.json                · the tokenizer, ~3 MB
├── tokenizer.json
└── tokenizer_config.json

Source repositories: argmaxinc/whisperkit-coreml (folder openai_whisper-small) and openai/whisper-small for the three tokenizer files.

Summaries and Ask AI: the language model (~2 GB)

Language models live in the shared Hugging Face cache, in a layout with internal links. Copy the whole model folder as an archive.

cd ~/.cache/huggingface/hub
tar -czf ~/Desktop/gemma-e2b.tgz models--mlx-community--gemma-4-e2b-it-4bit

On the restricted Mac:

mkdir -p ~/.cache/huggingface/hub
tar -xzf ~/Desktop/gemma-e2b.tgz -C ~/.cache/huggingface/hub

~/.cache/huggingface/hub/models--mlx-community--gemma-4-e2b-it-4bit/
├── blobs/         · the actual weights
├── refs/
└── snapshots/<id>/
    ├── model.safetensors
    ├── config.json
    ├── tokenizer.json
    └── …

If you would rather fetch it directly on the unrestricted machine, with the Hugging Face CLI installed:

pip install -U "huggingface_hub[cli]"
hf download mlx-community/gemma-4-e2b-it-4bit

That writes into exactly the folder above, ready to be archived and moved.

Check that Tape sees the files

1. Open Tape → Settings → Models.
2. Press the refresh button in the Speech Models or Language Models header.
3. The model you installed should now show its size on disk instead of a Download button.
4. For a language model, pick it under Summarization → Model. Only models that are actually on the Mac appear in that list, so its presence there is the real confirmation.
5. Record ten seconds of anything and stop. A transcript that appears with no network is proof it works.

Then turn on offline mode

With the models in place, switch on Settings → General → Work offline. Tape then makes no network requests at all: no update check, no model catalog, no downloads. On a filtered network that matters, because a blocked request usually hangs rather than failing, and a hang looks like a frozen app. With the switch on, anything missing is reported immediately and clearly instead.

Tape also offers this by itself. If a download fails the way a firewall makes things fail, the Models tab shows a banner with a Work Offline button and a link back to this page.

If a model shows up as damaged

An interrupted download leaves a truncated weight file behind. Tape detects that and offers Re-download, which clears the folder first. If you are installing by hand, delete the model folder and extract the archive again.

If macOS refuses to open the files

Files that arrived through a browser or an email client can carry a quarantine flag. Clear it on the folder you extracted:

xattr -dr com.apple.quarantine ~/Documents/huggingface/models
xattr -dr com.apple.quarantine ~/Library/Application\ Support/FluidAudio

Where everything lives, in one place

Settings → Models → Storage shows the total size and has a Show in Finder button that opens the models folder directly.