Which local model your Mac can run
An Apple Silicon Mac can transcribe a meeting and write the notes without anything leaving the machine. Installed memory decides which model does the writing.
An Apple Silicon Mac can transcribe a meeting and write the notes on-device. Which local LLM writes the notes depends on installed memory: under 16 GB, Qwen3.5 2B; under 24 GB, Qwen3.5 4B; under 32 GB, Qwen3.5 9B; at 32 GB or more, Qwen3.8 27B. Among the models that fit, pick by quality.
Which model can your Mac run?
Seven models from the mlx-community collection are offered, all 4-bit quantised, which shrinks the download and the memory they need at some cost in accuracy.
| Model | Download | Minimum RAM | Rating |
|---|---|---|---|
| Qwen3.5 2B | ~1.7 GB | 8 GB | 4 |
| Qwen3.5 4B | ~3.1 GB | 16 GB | 6 |
| Gemma 4 E4B | ~5.2 GB | 16 GB | 6 |
| Qwen3.5 9B | ~6.0 GB | 24 GB | 7 |
| Gemma 4 12B | ~6.8 GB | 24 GB | 7 |
| Gemma 4 26B (MoE) | ~15.4 GB | 32 GB | 8 |
| Qwen3.8 27B | ~16.1 GB | 32 GB | 9 |
The rating is an editorial 1 to 9 score, not a benchmark. Read it as an order.
The recommendation favours note quality over speed, which is why it points at the dense 27B at 32 GB. The Gemma 4 26B mixture-of-experts model is much faster and writes notes between the 12B and the 27B. Any model that fits is available, and a smaller one is a reasonable choice.
What each notes provider gets you
Every provider receives the same prompt, and notes stream into the window as they are written. What changes is where the prompt goes and who pays.
A local model keeps everything on your Mac and works offline. It is on the free plan. One model is loaded per generation and unloaded afterwards, so it does not hold memory between meetings. It is also the only provider that can run “Names from what people say”, the optional pass that names speakers from the transcript.
Apple Intelligence runs on-device on macOS 26 or later and downloads nothing. Its session budget is 4,096 tokens in total, so a long meeting is summarised in pieces and then condensed. A downloaded model handles long meetings more directly.
Your own Anthropic or OpenAI key sends the prompt from your Mac to that vendor, with no Listenr server in between. Also on the free plan.
Listenr Cloud is the default and needs Pro. What it sees is in what actually leaves your Mac.
Safeguards
- A model too big for your Mac is refused before the download, not after.
- Notes are never saved half written. A generation that hits the 8,192-token output limit counts as an error.
- A transcriber that fails to start does not cost you the meeting. Listenr falls back to Apple’s engine and keeps the audio, so you can transcribe the meeting again later.
Which transcriber, by language
Transcription models are small and stay loaded for the whole meeting. Pick by the languages you hear.
| Transcriber | Languages | Download |
|---|---|---|
| Apple (default) | One language you set | None |
| Parakeet Ultra (recommended) | 25 European languages, detected automatically | ~632 MB |
| Parakeet v3 | 25 European languages, detected automatically | ~483 MB |
| Parakeet v2 | English only | ~464 MB |
| SenseVoice Small | 50+, including Chinese, Japanese and Korean | ~500 MB |
| Whisper (whisper.cpp) | 99, detected automatically | 78 MB to 3.1 GB, default base |
| OpenAI (cloud, your key) | About 100 | None, audio is sent to OpenAI |
Parakeet and SenseVoice run on the Neural Engine. Parakeet Ultra is Moondream’s further training of NVIDIA’s Parakeet TDT v3. Whisper needs brew install whisper-cpp, which Listenr does not ship. Language counts come from the model makers, not from Listenr’s own measurements.
Speaker separation uses a separate on-device model, NVIDIA’s Nemotron diarization (about 103 MB), and a 1 MB voice-activity model (Silero VAD) splits the audio into windows.
Limits
Transcripts arrive a window at a time, not word by word. Local models are text-only, so attached screenshots reach the notes only through providers that accept images. Running on-device decides where your audio goes; it does not make the transcript more accurate than a cloud service, and Listenr publishes no benchmark either way.
What to install
For notes, pick the model your RAM supports and no larger. For transcription, try the default first, then move to Parakeet Ultra for a European language with automatic detection, SenseVoice for Chinese, Japanese or Korean, or Whisper for the widest language list. Each model downloads once, and after that a meeting needs no network.