You just finished a client meeting. You open Wispr Flow to transcribe your notes. A paywall pops up: $12/month. You can afford it — that’s not the point. The point is that your voice, your words, your data has to travel to someone else’s server before it becomes text on your screen. That quiet frustration? That’s exactly why FluidVoice exists.
It’s an open-source macOS dictation app that crossed 5,000 GitHub Stars by July 2026. It does what Wispr Flow and SuperWhisper do — you talk, it types — except everything runs locally on your Mac. No cloud. No subscription. No bytes leaving your machine.
The real battle isn’t Wispr Flow vs. FluidVoice. It’s cloud inference vs. local inference — and local is winning faster than anyone expected.
Here’s what makes FluidVoice different from the dozen other dictation tools cluttering the macOS landscape. Most commercial apps give you one model and say “take it or leave it.” FluidVoice hands you eight. You want near-zero latency? Switch to Parakeet Flash — 250MB, English-only, blazing fast on Apple Silicon. Need accuracy across 14 languages? Load Cohere Transcribe at 1.4GB. Running an older Intel Mac? There’s a Whisper build for you. This kind of model freedom is usually locked behind a paywall. FluidVoice gives it away.
Then there’s Fluid Intelligence — a 3.5GB local AI runtime that takes your raw transcription and intelligently formats it. Say “schedule a meeting at three pm tomorrow” and it outputs “Schedule a meeting at 3:00 PM tomorrow.” Capitalization, punctuation, number formatting — all handled on-device. Your voice data never touches a server.
But here’s the twist that has the open-source community arguing.
The core dictation engine is fully open-source under GPLv3. The enhancement layer — the thing that actually makes it feel magical — is closed-source. Free, but not open. Community ideals meet the brutal economics of funding development.
The team’s reasoning is honest: the free, open core gets adoption; the proprietary enhancement layer keeps the lights on. They’ve hinted it may open-source eventually. But right now, if you’re a developer who wants to audit the full pipeline, you hit a wall. If you’re a privacy purist who needs end-to-end verifiable security, the closed enhancement layer is a blind spot. And if the team ever pivots Fluid Intelligence to paid, the free version suddenly feels incomplete.
That tension — between openness and sustainability — is the most interesting thing about this project. It’s not a bug. It’s the future of open-source AI tools staring us in the face.
Installation is dead simple: one Homebrew command. brew install --cask fluidvoice. Grant microphone and accessibility permissions, pick a model, set a hotkey. The onboarding flow was specifically redesigned in v1.6.0 and it shows — clean, no jargon, each step explained in one sentence. For Apple Silicon users, start with Parakeet TDT v3: 500MB, 25 languages, the best balance of speed and accuracy. English-only users should grab Parakeet Flash for that “did it just read my mind?” latency.
Press your hotkey. Speak. Release. Text appears at your cursor. That’s it.
FluidVoice also splits dictation into two modes that most competitors don’t even think about. Write Mode drops text into any field — or rewrites existing content. Command Mode lets your voice control macOS: launch apps, run Shortcuts, trigger system actions. You can assign different prompt sets to different apps, so your formatting expectations for emails differ from code comments. This isn’t a voice input method. It’s a local-first voice workflow infrastructure.
On Reddit, someone wrote: “I cancelled my Wispr Flow subscription 2 days after installing FluidVoice. The Parakeet Flash model is faster than Wispr on my M2.” That comment didn’t get many upvotes. But it points to something bigger than a product comparison.
Local inference on consumer hardware has crossed the line where it can rival cloud services. The SaaS moat for voice transcription just developed a crack — and it’s widening.
The risks are real, though. The core team is 2-3 people maintaining 8 models across a planned 4-platform roadmap. That’s ambitious to the point of fragility. The bus factor is low. iOS, Windows, and Linux versions are promised but undated. The GPLv3 license switch (from Apache 2.0) raised eyebrows in the community — if you’re considering commercial integration, the copyleft implications matter. The Chinese language enhancement still lags behind English. Multi-speaker transcription isn’t supported — use Otter.ai or Feishu Minutes for that.
But the trend line is unmistakable. Local AI inference is the clearest technical direction of 2026. LLMs, TTS, STT — everything that can run locally is moving there. FluidVoice sits squarely on this wave, and in the macOS open-source dictation space, it currently has no equal. SuperWhisper and Ottex are local but closed and paid. FluidVoice is the only player combining open-source, free, and local — and that window won’t stay open forever.
So should you install it? If you have an Apple Silicon Mac on macOS 15, yes. Today. Run the Homebrew command, pick Parakeet Flash, set a hotkey, and use pure dictation for a few days before downloading Fluid Intelligence. Feel the latency for yourself.
The moment text appears on your screen the instant you stop speaking — with nothing traveling to a server, no subscription page loading, no meter counting your words — you’ll understand why cloud dictation’s days are numbered.
That split-second of zero latency is more honest than any review, any comparison table, any feature list. It’s the feeling of controlling your own voice again. And once you feel it, you won’t go back to paying $12/month for the privilege of using your own words.
FAQ
Q: If the enhancement layer is closed-source, isn't the 'open-source' claim misleading?
A: Partially. The core STT engine is genuinely GPLv3 open-source — you can audit, fork, and build on it. But Fluid Intelligence, the formatting and enhancement layer that makes output feel polished, is proprietary. It's a hybrid model: open foundation, closed magic. Whether that's acceptable depends on whether you need full-chain auditability or just want a fast, free dictation tool.
Q: Can a 2-3 person team actually sustain this long-term?
A: Honestly, the odds are against it at current scale. They're maintaining 8 models with plans for 4 platforms. The macOS version is solid, but iOS/Windows/Linux will stretch them thin. The project's survival depends on either expanding the maintainer team or the community contributing meaningfully. Right now the bus factor is dangerously low.
Q: Is local AI really going to kill cloud-based SaaS dictation?
A: Not kill — disrupt. Cloud services still win on multi-speaker transcription, cross-platform sync, and zero-setup convenience. But for single-user dictation on Apple Silicon, local inference has already matched or beaten cloud latency. The SaaS moat is shrinking. The services that survive will be the ones offering things local models can't — not basic transcription.