Changelog

Everything that shipped.

44 changes across 9 weeks of building, from the first build that put text at a cursor to version 1.0. Dates are the weeks the work actually landed.

  1. 31 August

    Every claim on this site, replayed against the app until each one was verified or fixed.

    • Better The whole public command reference now runs as a test: every spoken command, every alternative phrasing, and all 115 emoji names are replayed through the real pipeline before a release. The commands page is a test report, not a brochure.
    • Fixed "Star emoji" in a terminal produced "* emoji", because the symbol command fired before the emoji could. The two passes now know about each other, and "literally fire emoji" writes the words again instead of the flame.
    • Fixed Saying "literally" as an ordinary word ("it was literally amazing") no longer vanishes in terminals. It only acts as an escape when a command actually follows.
    • Better "cc the design team" is written CC, by fixed rule, while "250 cc engine" and "2 cc of saline" stay lowercase.
    • Better A word the engine mangles now repairs from your dictionary deterministically: "pymium" becomes premium because it is not a word and premium is yours. Real words never repair, so "commands" is never rewritten into "comments".
    • Better Terse note style survives polishing: "affect brighter" stays a fragment instead of being expanded into a sentence. Notes people write in shorthand come back in shorthand.
    • Better The lost-word guard now also catches paraphrases: output that trades "straight away" for "immediately" is discarded in favor of your exact words, while a genuine respelling of a mangled word still passes.
  2. 17 August

    The formatting week, part two: the vocabulary of work, and a formatter that can no longer lose a word.

    • Better Work vocabulary is written the way work writes it: "B to B" becomes B2B, "a b test" becomes A/B test, "q one" becomes Q1, plus 24/7, 10x, one-on-one, and around 200 acronyms and product names such as API, ROI, SaaS, iOS, LinkedIn and GitHub. Fixed rules, so it is right every time, and "from plan B to C" stays a range.
    • Better Spell a term correctly once and Flit repairs the copies the engine mangled: "me to be apps and B2B companies" comes out as B2B apps and B2B companies.
    • Better The formatting model can no longer quietly lose a word. Every content word and number you say is traced into the output, and a result that drops one is discarded in favor of your exact words.
    • Better With context awareness on, names on your screen become guaranteed spellings rather than suggestions, so a reply to Sarah Kwon spells her name the way the thread does.
    • New An optional larger formatting model for Macs with 32 GB of memory or more, for better judgment on messy dictation. The default stays light, and nothing downloads unless you ask.
    • Better The formatting test suite now replays batches of real dictations, transcription mistakes and all, so every change is measured against the speech people actually produce.
  3. 10 August

    One fix, and the right one: memory.

    • Fixed Flit no longer grows by gigabytes over a long day of dictation. A cache in the local model runtime was unbounded; it is now capped, and the app settles back to its resting size after every dictation.
  4. 3 August

    Every language Whisper can handle, and the routing fixes that made them usable.

    • New All 99 languages are now available, up from 46. 25 of them still stream live as you speak; the rest run on a slower local model.
    • Fixed Choosing "Auto" no longer hands the word "auto" to the speech engine as though it were a language.
    • Fixed Dictation now routes to the right engine even when automatic detection is switched off.
  5. 27 July

    Small edges, sanded.

    • Better Text around an expanded snippet is formatted with it, so a signature no longer lands mid-sentence with the capitals wrong.
    • Better Settings lost a redundant status label that repeated what the buttons beneath it already said.
  6. 21 July

    The formatting week. Most of what makes Flit write like a person rather than a transcriber landed here.

    • New The polish dial replaces the old on-or-off formatting switch, so you choose how much shaping you want rather than all or nothing.
    • New Flit learns the style you want per app from your own edits, instead of waiting to be configured.
    • New A dictation trace records why each result came out the way it did, shown as chips you can actually read and folded away until you want it.
    • New Formatting categories for notes, issues, social posts and search boxes, with the category each app gets shown in Settings.
    • New The recording HUD shows elapsed time and a stop button.
    • Better Formatting now runs for every language rather than English alone, using each language’s own punctuation conventions.
    • Better Spoken self-corrections are resolved without waking the model, so "no wait, Tuesday" costs nothing.
    • Better Spoken structure becomes written structure: lists stay lists, headings stay headings.
    • Better The prompt is assembled while you are still speaking, which removed the latency a richer prompt had cost.
    • Better Your dictionary now repairs confident-sounding nonsense, not just words the engine flagged as uncertain.
    • Better The menu bar panel opens anchored to its icon instead of drifting to a screen corner.
    • Fixed Bare times, pauses that should be commas, and dropped articles are handled by fixed rules rather than model judgement.
    • Fixed The model can no longer invent numbers that were never spoken, and units are normalised on the way out.
    • Fixed A transcript that comes back in the wrong script is repaired in the language you actually spoke.
    • Privacy Reading screen context is now opt-in, off until you turn it on.
    • Privacy Editing selected text by voice is opt-in too.
    • Privacy Screen context is kept out of the cached prompt prefix, so it cannot outlive the single dictation it was read for.
  7. 17 July

    v1.0

    The polish pass, and the look the app still wears.

    • New Orchid Sunset: the dark glass design now used across the whole app.
    • Better Model stack refresh, bringing a newer formatting model and a newer Whisper build.
    • Better A full polish audit across every screen, which is the point the app became something we would charge for.
  8. 11 July

    Everything needed to hand the app to somebody else.

    • New A downloadable formatting model for Macs without Apple Intelligence, so the writing quality does not depend on your macOS version.
    • New First-run onboarding that explains each permission before macOS asks for it.
    • Better An accessibility fallback for inserting text into Electron apps that refuse the normal path.
    • New The hummingbird: a proper app icon and menu bar mark.
  9. 17 June

    The first build that turned speech into text at the cursor.

    • New Hold a key, speak, and text appears where you were already typing. Everything else since has been making that sentence true in more places.

Updates arrive in the app. Your purchase includes 12 months of them, and the version you have keeps working whether or not you ever renew.

Your voice was always faster.

Not out yet. $29 USD once when it is, with no account, no subscription and nothing uploaded.

No spam, no sharing. One email when Flit ships, and you can leave in one click.

macOS 14 or later · Apple silicon