← Docs

On-device · no cloud

On-device cleanup

Parleq can run the cleanup pass entirely on your Mac — no API key, no cloud account, no per-call cost, and no network boundary for your transcript. There are two on-device options, both available under Settings → Cleanup → Provider — and the first-run setup wizard offers the on-device LLM directly as its On-device choice, with a picker for which model. Both keep your transcript on your machine; they differ in what they can do and what they cost in download and memory.

Two ways to run on-device

Both keep your transcript on your Mac — pick the track that fits, and switch anytime.

The Lightweight option — built on Concord (by Keavi) — is a fast, fully deterministic corrector that runs in-process, with nothing to download and almost no memory overhead. It handles the structured cleanup: adding the "?" when you dictate a question, writing the punctuation you speak (say "comma", "period", "question mark", or "open/close quote" and it becomes the mark), spelled-out numbers and percents (so "forty five percent" becomes "45%"), joining split compound words, your custom-dictionary terms — matched by how they actually sound (phonetic distance over dictionary pronunciations, not a confidence guess), plus spoken forms — and, once you've enrolled them, voiceprint disambiguation of sound-alike words. General sentence casing, and the punctuation the speech model infers on its own, still come from the model itself — this tier adds the marks you speak explicitly. Because there is no language model on this path, it can't freely rewrite text, so the voice-refine loop, the one-tap style chips, and per-app styling aren't available here; switch to the on-device LLM or a cloud provider when you want those.

To use it, open Settings → Cleanup → Provider and select Lightweight (on-device). (It's a Settings option — the first-run wizard's On-device choice is the on-device LLM below, with a model picker.) There's nothing to download and nothing to configure — it's ready immediately, and like every on-device path your transcript never crosses a network boundary.

The engine, up close

A cluster of small, fast corrections.

On-device cleanup isn't a black box — it's a handful of deterministic helpers, each doing one thing well, all running in-process on your Mac. Built on Concord (by Keavi).

  • Voiceprints
  • Questions
  • Numbers
  • Dictionary
  • Compound

The on-device LLM is for fuller cleanup and the full voice-refine loop — pick it when transcript content must never leave your machine but you still want a model to rewrite, not just tidy. There's a choice of two models: Gemma 4 E4B, the default and the higher-quality of the two, and Qwen3-4B, a lighter model that brings on-device LLM cleanup to Macs with as little as 8 GB of RAM.

At a glance

  • Auth: None. No account, no key.
  • Models: Both 4-bit, run via MLX, and downloaded once from Hugging Face:
    • Gemma 4 E4B (default) — ~4 GB download, ~6 GB while loaded, needs a Mac with at least 12 GB of RAM.
    • Qwen3-4B (lighter) — ~2.4 GB download, ~4 GB while loaded, needs a Mac with at least 8 GB of RAM.
  • Memory: Only while the model is loaded; Parleq frees it after a few minutes of inactivity.
  • Requirement: Whichever model needs more RAM than your Mac has, that model appears in the picker but can't be selected — the other one still can.
  • Network: None for cleanup — your transcript never crosses a network boundary. The only network use is the one-time model download you initiate.
  • Cost: $0. No per-call charges, ever.

Setup

  1. 1.

    Choose On-device, then a model.

    On first launch the setup wizard asks how you want cleanup handled. Pick On-device, then choose Gemma 4 E4B (the default, best quality) or Qwen3-4B (lighter, for lower-RAM Macs). Already past the wizard? Open Settings… → Cleanup → Provider and choose On-device (no cloud) — the same option, labeled there to set it apart from Lightweight — then pick your model.

    Gemma 4 E4B needs a Mac with at least 12 GB of RAM; Qwen3-4B needs at least 8 GB. On a Mac below a model's RAM floor, that model is visible in the picker but disabled — its working set would crowd out your other apps, so Parleq won't let you select it.

  2. 2.

    Let the model download.

    Selecting a model starts a one-time download from Hugging Face — about 4 GB for Gemma 4 E4B, or about 2.4 GB for Qwen3-4B. Real, byte-accurate progress shows in the wizard and in the menu-bar item, and the download resumes where it left off if it's interrupted — quit, lost network, even a force-quit — rather than starting over. You can keep using Parleq while it downloads (it pastes raw transcription until the model is ready). The model is stored at ~/Library/Application Support/Parleq/models/.

  3. 3.

    Test a dictation.

    Hold right Option, say a sentence, release. The first dictation after launch loads the model (a couple of seconds); after that, cleanup runs in well under a second, entirely on your Mac.

Memory & residency

While the model is loaded it uses about 6 GB of memory for Gemma 4 E4B, or about 4 GB for Qwen3-4B. Parleq keeps it resident only as long as you're actively dictating and unloads it after a few minutes of inactivity, returning that memory to the system. The first dictation after an unload reloads the model (a couple of seconds); subsequent ones are fast.

Parleq also caps the pool of reusable GPU scratch buffers the model keeps around, so the memory it holds while loaded stays close to the model's own size rather than climbing several gigabytes higher after a longer cleanup. The figure you see in Activity Monitor therefore tracks the model itself — useful to know if you or your security team are watching the process's footprint.

You can change this in Settings → Cleanup with the Keep model loaded control — Automatic (the default), Keep loaded (no reload pause, but holds the memory), or Unload when idle.

Removing the model

To reclaim the disk space, open Settings → Cleanup and click Remove downloaded model. That deletes the weights for whichever model you downloaded from ~/Library/Application Support/Parleq/models/ and frees the in-memory copy immediately. You can re-download anytime by selecting On-device (no cloud) again, switch to the other model, or switch to Lightweight, a cloud provider, or no cleanup at all.

Privacy & licensing

Nothing about your dictation leaves your Mac on this path. Audio stays on-device (as always), and with on-device cleanup the transcript never crosses a network boundary either — there is no provider receiving your text. See the privacy policy for the full data posture.

Both models are Apache 2.0 licensed: Google's Gemma 4, under its Apache 2.0 license, and Alibaba's Qwen3-4B, also Apache 2.0. Parleq doesn't redistribute either model's weights — your Mac downloads them from Hugging Face when you enable this option.