Most dictation tools, even local ones, use Whisper or similar offline models: you record, then wait for the transcript. Localvoxtral uses Mistral's Voxtral Realtime, one of the first open-source speech models with a natively streaming architecture. Words appear as you speak, not after you stop. It feels closer to someone typing along as you talk.

Press a shortcut, speak, and text gets typed directly into whatever app you're in. No cloud, no subscription, no data leaving your machine.

Two backend options:

voxmlx on Apple Silicon: I forked voxmlx to add a WebSocket server and memory optimizations. Runs a 4-bit quantized model on an M1 Pro. Audio and inference stay fully on-device. vLLM on NVIDIA GPU: tested on an RTX 3090, noticeably faster.

The app is native Swift (~97%), lives in the menu bar, and stays out of your way. Configurable shortcut, mic selection, auto-paste. GitHub: https://github.com/T0mSIlver/localvoxtral

Pre-built DMG available in Releases