FluidVoice: on-device macOS dictation with a private AI enhancement layer

What FluidVoice solves, and who it is actually for

Most dictation tools send audio to a server. FluidVoice takes the opposite position: speech recognition runs on the Mac, and the optional post-processing layer, Fluid Intelligence, also runs locally. The README frames this as "No cloud, no API keys, no data leaving your Mac." That is the whole pitch, and it is a narrow one. The project is for macOS users who dictate into other applications all day and who do not want their audio or transcripts crossing the network.

The README is explicit that FluidVoice is "on macOS today" and that iOS and Windows are "on the way," with a waitlist at altic.dev/fluid/waitlist. So the audience is narrower than the feature list suggests. If you are on Linux or Android, there is nothing here for you yet, and the repository contains no build target for those platforms.

The second audience is people who want voice to do more than insert text. Command Mode lets you launch apps, run shortcuts and trigger system actions by voice. Write Mode rewrites selected text in any text field. Those two modes are the reason someone would pick this over a plain transcription utility.

The split architecture: open app, closed enhancement runtime

This is the most important structural fact about the project, and it is easy to miss. The application is open source under GPL-3.0. Fluid Intelligence is not. The README says it plainly: "Fluid Intelligence is a separate, privately maintained local AI runtime that powers advanced on-device dictation enhancement." The stated reason is sustainability, so the core dictation experience can stay free, and the README adds that this "may change in the future."

That means the repository you can clone does not contain everything the shipped app runs. The build produces the dictation app, the speech model integrations, the overlay, the hotkey handling and the settings surface. The enhancement model is delivered separately, and the README points to X (@fluidvoiceapp) for "exclusive model access." Anyone evaluating the project on licence grounds should treat these as two products with two different distribution models.

On the speech side the app is more open. The supported models listed in the README are Nemotron Speech 3.5, Parakeet Flash, Parakeet TDT v3 and v2, Cohere Transcribe, Apple Speech, and Whisper. The README links the Parakeet realtime model to a Hugging Face page under nvidia, which tells you where those weights come from. You can run FluidVoice with any of them and skip Fluid Intelligence entirely.

Installing FluidVoice with Homebrew and dictating once

The README gives two install paths: a Homebrew cask and a manual download of the latest release. The cask is one line.

brew install --cask fluidvoice

After that, launch the app from Applications. The 1.6.0 release notes describe a "language-first voice engine setup" in onboarding, followed by a real dictation tryout and AI enhancement setup. Expect to grant microphone access and accessibility permission, because the README says text insertion happens "via accessibility APIs" for app-independent entry.

The global hotkey is the main interaction. Press it, speak, and the transcription appears. The README describes a Live Preview overlay with notch support, so on a notched MacBook the words appear around the notch as you talk.

If you want the app to work without any cloud provider, leave the AI enhancement step unset during onboarding. The README lists OpenAI, Groq and custom providers as optional post-processing, alongside the local Fluid Intelligence option. Skipping all of them still gives you local speech recognition.

For a manual install instead of Homebrew, the README points at the releases page:

open https://github.com/altic-dev/FluidVoice/releases/latest

Where FluidVoice stops being the right tool

The platform boundary is the first hard limit. No Windows build is published in this repository, no Linux build, and no iOS build. The README says iOS and Windows are coming and directs readers to a waitlist. If your team is not on macOS, the evaluation ends there.

The second limit is the closed enhancement runtime. If your reason for choosing an open source dictation app is auditability of every component that touches your audio, FluidVoice only partly satisfies that. You can read the app, but you cannot read Fluid Intelligence. The README does not document what the runtime does with audio beyond describing smart formatting, context-aware capitalization and post-processing, and it does not describe a rollback path if an enhancement mangles a transcript.

The third limit is permission fragility. Insertion depends on accessibility APIs, which macOS gates per application. The README does not enumerate which apps are tested. A password field, a terminal multiplexer or an app with custom text handling is a plausible place for insertion to behave differently, and the documentation is silent on that.

Finally, the README's own warning about Fluid Intelligence is marketing, not engineering: "Fluid Intelligence may cause you to unsubscribe from other dictation apps and save money." Treat that as a claim to test, not a specification.

FluidVoice against Wispr Flow and Handy

The README describes FluidVoice as "A local Wispr Flow alternative," so the comparison is fair game. Wispr Flow is a cloud dictation product: audio leaves the machine for transcription. FluidVoice's core claim is the inverse, with recognition on device and an optional local enhancement layer. That difference is not about accuracy, it is about where the audio goes and whether the tool works without a network connection.

Against Handy, another name people search for alongside FluidVoice, the distinction is scope rather than privacy. FluidVoice bundles Command Mode and Write Mode, plus per-app prompt sets, so the same hotkey can behave differently depending on the application in focus. A minimal transcription utility does one thing: it turns speech into text at the cursor. If that is all you want, the extra modes are surface area you will not use.

The honest comparison point is model choice. FluidVoice exposes Nemotron Speech 3.5, Parakeet Flash, Parakeet TDT v3 and v2, Cohere Transcribe, Apple Speech and Whisper, and the README claims "one of the fastest native implementations of Parakeet." That claim is unverified here. What is verifiable is that the model list is broad and that Apple Speech gives you a zero-download option on macOS.

Maintenance cadence, GPL-3.0 obligations and upgrade cost

The repository is not archived, and the last push was on 2026-09-15. Releases have been frequent: v1.6.7 on 2026-08-05, v1.6.8 on 2026-08-11, and v1.6.9 on 2026-08-18. That cadence matters for a dictation tool, because macOS changes break accessibility hooks and audio capture regularly.

The licence is GPL-3.0. For individual users this is unremarkable. For a company that wants to embed FluidVoice in a product, or ship a modified build, GPL-3.0 carries source disclosure obligations that a permissive licence would not. The README itself notes the project is "free and open source under GPLv3" and asks for stars to keep development going. This is not legal advice; if you plan to redistribute anything built on this code, get your own counsel.

Upgrade cost is low by design. The README lists auto-updates with an optional beta channel, and the Homebrew cask means `brew upgrade` keeps the app current. The real upgrade risk is the Fluid Intelligence runtime, which ships separately from the open repository. A change there is not visible in the commit history you can read.

Editorial conclusion

Adopt FluidVoice if you dictate on macOS and want speech recognition and optional enhancement to stay on the machine, and if you accept that the enhancement runtime is closed and the licence is GPL-3.0. Do not adopt it if you need Windows, Linux or iOS today, or if you cannot ship GPL code. Verify two things first: whether the Homebrew cask resolves on your macOS version, and whether the app's accessibility and microphone permissions work in the specific apps you dictate into.

Frequently asked questions

Official sources

1. altic-dev/FluidVoice on GitHub
2. License: GPL-3.0
3. Project website
4. README
5. Releases