Latest in AI
Hex 3.0: Local AI Dictation for Mac, Windows and Linux
Anomaly's voice tool has a new stable release, but existing users should read the migration notes before replacing their installation.
Anomaly released Hex 3.0 on October 31, bringing an updated local voice-dictation app to Mac, Windows and Linux. It adds vocabulary controls and a shared settings interface, while changing how existing installations upgrade. Ordinary dictation runs on your device rather than requiring a cloud transcription service.
The OpenCode announcement positions Hex as a way to speak to coding agents. Its scope is broader: the app transcribes speech and pastes text into the application you are using. OpenCode is optional.
What changed in Hex 3.0
The versioned release notes describe a common settings window across the three operating systems, a new recording indicator and vocabulary entries for names and specialist terms.
Linux gains ARM64 builds and Debian and RPM packages.
Meetings remains a beta feature in the macOS app. Annotate, which combines screen recording, drawing and speech for an OpenCode session, belongs to the Dev channel on macOS and Linux.
Upgrading from Hex 2.3.1 needs a fresh setup
Version 2.3.1 does not automatically update to 3.0. Users must download the replacement, which starts with an empty data folder rather than importing settings, history, meeting recordings or downloaded models. The old folder remains untouched, and previously exported meeting notes remain where they were saved.
Mac users must grant permissions again; Windows users must remove the old installation separately.
Check what you need from the old installation before replacing your working setup.
Local transcription does not mean every optional feature stays local
Hex's repository documentation distinguishes on-device speech recognition from optional OpenCode rewriting. Ordinary dictation needs a downloaded speech model, but does not need OpenCode or a cloud transcription provider.
Optional rewriting sends the transcript and application context to the configured model provider. That is a different data path from local transcription. A team evaluating Hex for sensitive work should decide whether it wants transcription alone or provider-based rewriting as well.
The repository uses the MIT license. The download page provides desktop installers; the Mac build requires Apple silicon and macOS 15 or newer. Windows and Linux have their own platform requirements, and the repository still describes those ports as beta.
Why a separate dictation layer matters
A voice feature inside one coding tool ends when you switch applications. A system-level dictation tool can follow the same workflow into an editor, issue tracker or document without requiring each product to implement speech input.
That makes portability the useful distinction here. The thing to evaluate is whether spoken instructions reach the intended field accurately and consistently, including names and technical terms. A local model changes where transcription happens; it does not remove the need to review what gets pasted.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help technical founders turn product launches and implementation details into clear stories developers can understand and share.