VoxtMini

VoxtMini Changelog

The apps check for updates once a day and show these notes before installing.

All apps' changelogs
Release notes languageEnglish简体中文
  1. 0.1.0-beta.8Beta

    Fixed

    • Local models work on macOS 15 again: the speech runtime was built for macOS 26 only, so on macOS 15 every local speech model failed to start ("language version 4.0 which is not supported on this OS"). Releases are now checked so that nothing in them needs a newer macOS than 15.0. If this ever happens again, VoxtMini says that this build does not support the Mac's macOS version.
  2. 0.1.0-beta.7Beta

    Added

    • Send Feedback: a window of its own to send a problem, a suggestion or a question to actnow.dev from the app, with screenshots or a screen recording (previewed in the window) and, if you want, diagnostics you can look at before sending (versions, models, settings, recent log lines and the last crash, without what you said or wrote, keys, names or addresses). Replies show in the app and by email; Settings → Feedback lists what you sent. Feedback written offline is sent once you are online. Help → Send Feedback, the menu bar and the command palette open it; GitHub issues still work.
    • Report this problem beside a failed dictation (History), a failed try, the helper's card, a model whose download failed and a history that could not be read, and the crash card's Send feedback: the feedback window opens with an error report (the error, where it happened, your last few actions by name and the log lines of that moment, without what you said or wrote), which you can leave out.
  3. 0.1.0-beta.6Beta

    Fixed

    • Fixed a crash when logs contained malformed links: a link VoxtMini could not read (such as one with a broken % escape) in what it writes to its log could make it quit. Such links are now logged safely, with any password or key in them still hidden.
  4. 0.1.0-beta.5Beta

    Added

    • Remote model services on your own network: an endpoint may use plain http:// (ws:// for streaming) to a LAN or Tailscale address (for example http://100.88.1.2:3425/v1) or a host name on it, not only to this Mac. A host name is checked when connecting: if it points to a public address, nothing is sent. The setup dialog notes that such a connection is not encrypted. If macOS blocks it, the error asks you to allow VoxtMini under System Settings → Privacy & Security → Local Network.
  5. 0.1.0-beta.4Beta

    Changed

    • The floating overlay changes size smoothly (live text growing, the answer card, a notice) instead of jumping; with Reduce motion on it changes at once.
    • While the microphone or a model is getting ready, the overlay shows a turning hourglass instead of a line of text.

    Fixed

    • The first dictation after VoxtMini opened could stay on "Starting the microphone" while the speech models were still loading, end with nothing, and play two sounds. Recording now starts at once; if the microphone has not started when you stop (or after 12 s), VoxtMini says so with one sound. While macOS asks for the microphone, the overlay says so.
    • A microphone that started after its dictation had already ended stayed in use until VoxtMini quit.
  6. 0.1.0-beta.3Beta

    Added

    • Remote models, by your choice: set up remote speech and language services on the Models page (OpenAI, Doubao, Alibaba Cloud, Gemini Live, StepFun, Z.ai, Xiaomi MiMo; Anthropic, Gemini, DeepSeek, OpenRouter and others; Ollama, LM Studio and oMLX on your Mac) and choose one on a feature's page. Only that feature sends its audio or text, and only to that service; "Automatic" is always a local model. Keys stay in the Keychain and can be shared with the other ActNow apps. OpenAI-style services can use the Responses API, and Rewrite can search the web.
    • Apple Speech and Apple Intelligence as local models, with nothing to download (on this Mac only; Apple Intelligence needs macOS 26).
    • One Models page for speech and language models, local and remote, with filters; models can be hidden. More local models, marked experimental: Whisper large-v3-turbo, large-v3 and small (no live text), Qwen3-ASR 1.7B 6-bit, Gemma 4, Qwen3.5 9B, Ministral 3, LFM2 and others.
    • Dictionary: categories, how often each term was used, terms limited to an App enhancement group, and the corrections it made shown in the history.
    • History: play an entry's recording; ask follow-up questions about an entry, answered by Rewrite's model and kept with it.
    • Shortcuts: mouse buttons, and presets (fn included); Space starts a follow-up while Rewrite's answer card shows.
    • Output and recording: keep the result on the clipboard (on by default), Escape cancels a recording (on), show live text (on), and a voice end command (off): say "over", 完毕 or 以上 and pause to stop.
    • Show icon in Dock (off by default); closing the window keeps VoxtMini running with either icon.
    • Notifications when a download finishes or fails while the window is not in front; a Microphone submenu in the menu bar; reorder microphones by dragging them.
    • Announcements from actnow.dev, shown once as notifications.
    • One editable prompt per feature, in place of "Your rules": Transcribe's clean-up, Translate and Rewrite each show their prompt, to change as you like, with variables to insert and Reset to default. The input, the dictionary, App enhancement's scene and the output format stay VoxtMini's (see "Show the full prompt"). Rules you had written are kept, added to the end of the prompt.
    • Settings per model, from a model's ⋯ menu on the Models page: only the knobs that model has (Thinking, temperature, top-p, top-k, min-p, repetition penalty, maximum output; a main-language hint for speech models); empty keeps VoxtMini's tuned values. Compact model cards.
    • A self-check of the network, disk and Keychain under Settings → About, with a proxy connection test; a crash report is offered at the next start.

    Changed

    • Languages: English, Chinese and Japanese. Chinese results are always in Simplified characters (any Traditional characters a model writes are converted); Japanese text is never altered. Models that do not know the three were removed.
    • Shorter text everywhere: explanations moved behind ⓘ; a smaller Settings window.
    • The Speech Recognition permission is asked for only when a feature uses Apple Speech.
    • VoxtMini now lives in the former Voxt repository on GitHub (hehehai/VoxtMini); Voxt's history ends at the tag voxt-final.
    • VoxtMini has its own page on actnow.dev (actnow.dev/voxtmini), linked from Settings → About with its download; updates come from its own feed there.
    • VoxtMini has its own identity on the Mac (bundle id dev.hehehai.voxtmini, folders ~/Library/…/VoxtMini, links voxtmini://), so it no longer shares one with Voxt 2. The first start moves your history, recordings, dictionary, settings and model claims over by itself. macOS asks once more for the microphone and Accessibility: grant them again.

    Fixed

    • A dictation into a web page's text box (seen in Arc) sometimes did not go in, and the answer card said "The text did not show up where you type" although the cursor was in the box. VoxtMini now sends its keys as Voxt does, after every app's keyboard hooks, so a busy Mac cannot hold them up; and when the box still shows nothing half a second after ⌘V, with nothing else changed, it sends ⌘V once more before showing the card.
    • Clicks anywhere on the Mac could stall for seconds when VoxtMini's helper was swapped out: the helper now only listens to the mouse.
    • Permissions no longer flicker when VoxtMini comes to the front.
    • Live text no longer stops after the speech model was slow to start.