State-of-the-art models, running on your silicon
Tape downloads a model once and then runs it on the Apple Neural Engine and GPU through Core ML and MLX. Nothing is sent to an API, so there is no per-minute cost and no meeting leaves the machine.
| Model | Job | Languages | Size |
|---|---|---|---|
| Parakeet TDT 0.6B default | Transcription, fast and accurate | English | ~600 MB |
| Whisper (tiny → large-v3) | Transcription, multilingual | English, Português and 97 more | 75 MB – 3 GB |
| Gemma 4 E2B default | Summaries and chat | 140+ | ~2 GB |
| Gemma 4 E4B | Summaries and chat, higher quality | 140+ | ~4 GB |
Choosing a speech model
The engine setting defaults to Automatic, which picks Parakeet for English and Whisper for everything else. Parakeet is dramatically faster; Whisper is the one that speaks other languages and can work out which language it is hearing.
If you set a language other than English, Tape uses Whisper even when Parakeet is selected. Forcing an English-only model onto a Portuguese meeting produces confident nonsense, which is worse than being slow.
Whisper sizes
Bigger is more accurate and slower. small is a good default for most meetings; large-v3 is worth it for heavy accents, poor microphones or several people talking over each other.
Language detection
Leave the language on Automatic and Whisper identifies the language before transcribing, rather than translating on the sly. The detected language is stored with the transcript, and the summary follows it unless you pin a language of your own.
Choosing a summarization model
Summaries and Ask AI run through MLX. Gemma 4 E2B is the default and comfortable on any Apple Silicon Mac. E4B writes noticeably better summaries and wants more memory. Only models actually present on your Mac are offered in the picker, so there is never a surprise multi-gigabyte download in the middle of a meeting.
You can also point Tape at other MLX models you have downloaded yourself. Anything that shows up in the shared model cache appears in the list.
How much of a long meeting the model reads
A two-hour meeting does not fit in a small model's context, so Tape keeps the opening and the ending and drops the middle. The slider in Settings controls that budget, in characters, and tells you how many minutes of speech it corresponds to and how much memory the context will need. It starts at whatever your Mac handles comfortably.
Where models are stored
Models live in shared caches outside the app, so reinstalling Tape never means downloading them again.
| Model | Folder |
|---|---|
| Parakeet | ~/Library/Application Support/FluidAudio/Models/ |
| Whisper | ~/Documents/huggingface/models/argmaxinc/whisperkit-coreml/ |
| Language models | ~/.cache/huggingface/hub/ |
Settings → Models lists everything on disk with its real size, and lets you download more or move old ones to the Trash. An interrupted download is detected and offered as a repair rather than failing later with something cryptic.
On a network that blocks model downloads? Corporate proxies and firewalls often block Hugging Face, where the models are hosted. You can still install them by hand: see installing models on a restricted network.
Credit where it is due
- Parakeet TDT by NVIDIA, in the Core ML port maintained by FluidInference
- Whisper by OpenAI, in the Core ML port maintained by Argmax as WhisperKit
- Gemma by Google, quantized for MLX by the mlx-community
Tape is an independent app and is not affiliated with any of them.
All of it, running on your Mac
No API keys, no per-minute billing, no account.
macOS 15+ · Apple Silicon