2.0.0

Benescript, now included

• Select text in any app, tap the trigger (Right Option by default), and the same AI presets rewrite it in place
• Magic sequences: type /// inside a selection to give an instruction, or //? to ask a question about it
• Quick slots (⌃⌘1-0) run a preset from the HUD, and any preset can take its own hotkey and run with no HUD at all
• Pin a separate AI provider and model for text enhancement, or follow the Viva Mode of the app the selection came from
• Shares presets, vocabulary and word replacements with dictation. Off by default; find it under Text Enhancement

Local AI

• MLX and Gemma are now one Local AI provider with a single model list, each card tagged with its runtime
• Gemma 4 E2B, E4B and 12B on LiteRT-LM; MLX moves to Qwen3.8 27B, Qwen3.6 35B-A3B, gpt-oss-20b and Qwen3.5 2B
• Two models are recommended for your specific Mac based on its memory
• Nothing loads at launch. The model warms when you start recording, and streams its output as it writes

Transcription

• Speaker labels, configured per Viva Mode, so a meetings mode can label speakers while a dictation mode stays clean
• Translate while transcribing on Soniox, Gladia and Speechmatics: no second AI pass, no extra latency, no extra cost
• Transcribe a queue of files instead of one at a time, each row with its own state and retry
• Nine new on-device models, including Parakeet 110M at around 100x realtime, Canary 1B v2 across 25 European languages, and Qwen3-ASR across 30
• Five new cloud providers: Gladia, Speechmatics, AssemblyAI, Cartesia and xAI Speech-to-Text
• Gemini 3.5 Transcribe and Deepgram Flux; transcription clean-up is now a toggle
• Filler removal follows the transcript language, so German and Portuguese no longer lose the real word "um"

AI providers

• Sign in with a SuperGrok or X Premium subscription and use Grok for AI processing and transcription without an API key
• Measured rankings ship in the app: every model graded on real dictation transcripts and timed, with a badge on those with a usable free tier
• New providers: MiniMax, OpenCode Zen, OpenCode Go and Ollama Cloud
• Gemini 3.7 Flash, Grok 4.6, GLM-5.3 and a full GitHub Copilot refresh. Retired models are mapped forward automatically

Notes and modes

• Share notes as Markdown through the standard share sheet, or have a mode auto-write every note into a folder
• A tag can shield notes from auto-delete, so cleanup can stay on for everything else
• Turn a note into Apple Reminders drafts, and optionally Calendar events, reviewed before anything is added
• Quick-add panel for vocabulary and word replacements behind a global shortcut
• A more generous free allowance: 15 uses a day during the trial, up from 10

Fixes

• Cloud uploads no longer negotiate HTTP/3, which some VPN tunnels blackholed until the request timed out
• Both local engines were misreading the start of every recording, transcribing 4 KB of padding as audio
• Line breaks survive filtering, so a transcript with speaker turns no longer arrives as one wall of text
• Cartesia, Cohere, Gladia, Speechmatics and AssemblyAI API keys can now be saved
• Your custom dictionary reaches Deepgram; xAI no longer forces English on auto-detect; Mistral honors the mode language
• A streaming session that returns nothing falls back to the recorded file instead of pasting an empty transcript