v0.4.1

With hold to talk turned off, tapping the shortcut to stop a dictation no longer starts a new one straight away. The island now shows the transcribing animation and leaves, as it does with hold to talk.

Flow 0.4.0

Cleanup now has two levels, none and light. A saved medium or hard setting keeps working, as light.

• Noise suppression is removed. Turning it on also turned off streaming transcription, adding seconds to every dictation, and it did not improve recognition. A config that still sets it is ignored.
• The last word of a dictation is no longer cut off.
• Your clipboard is kept when dictations queue back to back.
• Cleanup stays on Gemini 3.1 Flash-Lite, re-tested against newer models and still the most accurate and fastest.

Stable and experimental start from the same code in this release.

Flow 0.4.0 experimental

Experimental builds are previews for testing upcoming changes before they reach stable. They may change more often or contain unfinished features. Use Stable for everyday dictation.

This build starts from the same code as Flow 0.4.0 stable.

Flow 0.3.2

Modifier-only shortcuts such as Ctrl+Shift and Super+Shift now work for hold-to-talk.

• Consistent, more readable console fonts.
• Simpler settings copy and a cleaner vocabulary empty state.
• Opening the console reuses an already-listening daemon.

Stable and experimental start from the same code in this release.

Flow 0.3.2 experimental

Experimental builds are previews for testing upcoming changes before they reach stable. They may change more often or contain unfinished features. Use Stable for everyday dictation.

This build starts from the same code as Flow 0.3.2 stable.

Flow 0.3.1

Flow 0.3.1 brings streaming local dictation to stable, with Nemotron speech recognition on-device and Gemini 3.1 Flash-Lite cleanup through OpenRouter.

• Updates keep the daemon and console on the same channel, verify versions, and reopen the replacement window independently.
• The shortcut picker accepts single keys, single modifiers, and combinations; capture completes on release, cancellation is isolated, and failures show feedback.
• Model setup and repair stay in the console. Saved OpenRouter keys take effect without restarting.
• Standing cleanup instructions are removed, and Medium cleanup preserves vague references.
• Stable and experimental start from the same implementation. Experimental is an opt-in channel for upcoming changes, with generic descriptions in Settings.

Validated with workspace and console tests, lint checks, packaging regressions, local streaming tail/reset checks, and live input-device shortcut capture.

Flow 0.3.1 experimental

This experimental release starts from the same implementation as stable 0.3.1. Only the release version and channel metadata differ.

It includes the streaming dictation, update, setup, and shortcut fixes from stable 0.3.1. Future test changes will land on the experimental channel before stable.

Enable Experimental build in Settings to opt in. Disable it to return to stable.

Flow 0.3.0 — local stable with experimental opt-in

Flow 0.3.0 keeps dictation fully local with Parakeet and Qwen, and adds an opt-in experimental release channel.

• Open Settings → Build → Experimental build to download the MAI + Flash-Lite release, then click Restart Flow.
• Experimental sends audio and text through OpenRouter and requires your own API key and paid usage. Stable remains local.
• Turn the toggle off and restart to return to your saved stable build, including without an internet connection.
• Updates stay in the selected channel. Downloads are verified against the release checksum.
• Local models, settings and history are preserved when switching.

Install: unpack the Linux x86_64 archive and run bash packaging/install.sh, then open Flow. Existing users can update from About. New stable installs fetch their local models in the app.

Validation: daemon and console tests, clippy, GitHub release build, archive checksums, installer migration tests, bad-download/checksum rejection, and switching both packaged builds in an isolated install.

Flow 0.3.0 experimental — MAI + Flash-Lite

This opt-in release uses Microsoft MAI-Transcribe-2 for speech recognition and Google Gemini 3.1 Flash-Lite for cleanup, through OpenRouter.

1. Install or update to stable v0.3.0.
2. Enable Settings → Build → Experimental build and click Restart Flow.
3. Enter your OpenRouter API key in Settings, then start dictation.

Audio and text leave your device. OpenRouter usage is billed to your account; network access is required. Stable retains the local Parakeet + Qwen pipeline. Turn the experimental toggle off and restart to return to it. Both binaries, local models, settings and history are retained.

This package is marked experimental and cannot overwrite the stable binary. Updates follow the selected channel. The retired benchmark runner is not part of this release.

Validation: daemon and console tests, clippy, GitHub release build, archive checksum verification, failed-download/checksum protection, and stable → experimental → stable using the packaged binaries. Live provider behavior was not re-tested for this release because no OpenRouter key was configured in the test environment.

v0.2.0-alpha.2

What's Changed

• fix(console): offer the updates that actually exist by @Genoux in #13
• chore(release): 0.2.0-alpha.2 by @Genoux in #14

Full Changelog: v0.2.0-alpha.1...v0.2.0-alpha.2