Every small change, written down

• Improved the rewrite, summary, and translation prompts: rules are now unified and clearer, and translation keeps the original paragraphs and lists where possible.
• Switching to a cloud model, turning off rewriting, or deleting a local model now frees the memory the local model was using.

• The multilingual rewrite model is now Gemma 4 E4B, with better quality (re-download required).
• New installs use the Structure rewrite style by default.
• The vocabulary learns more eagerly and remembers more of your corrections.
• Model cards now say "Faster" or "Better quality" to make choosing easier.
• Sidebar navigation uses shorter labels, and Settings drops repeated intro text.
• Pressing the dictation key while the last sentence is still processing plays a short "please wait" sound.
• The microphone turns off as soon as recording ends.
• Quitting the app is faster.

• Fixed recording occasionally getting stuck until the app was restarted.
• Fixed recording occasionally not stopping after releasing the dictation key.
• Fixed rewriting staying off after the rewrite model quit unexpectedly.
• Fixed changing model settings mid-processing interrupting the current sentence.
• Fixed the sidebar memory figure showing far less than the real usage.
• Fixed the too-long-dictation message pointing to a setting that doesn't exist.
• Fixed the vocabulary occasionally learning a whole sentence after a full rewrite.
• Fixed the clipboard not being fully restored after copying multiple files.
• Fixed shortcut binding still recording keys after switching to another window.

• Language variants: in Settings → General, choose the written form of transcripts — US or UK spelling for English, Simplified or Traditional characters for Chinese.

• Chinese number conversion feels more natural: consecutive numeral readings now become Arabic digits by default (e.g. 「二零一八届」→ 2018届), while fixed expressions stay in Chinese characters.
• After switching or editing the polish model, the first sentence is no longer noticeably slow.
• History tab counts follow your search, and edited entries show an "Edited" badge.
• The "Use recommended" button on the models page now starts downloading the recommended model when it isn't installed yet.

Performance

• Cloud polish responds faster (measured median time from about 3.8 s to 1.7 s).

• Fixed the History page failing to open, and its counts and "Clear" button being inaccurate, after upgrading.
• Fixed dictionary auto-learning not working after upgrading.
• Fixed model downloads that stall and stay unresponsive for a long time.
• Fixed the whole sentence being lost when translation fails.
• Fixed English words occasionally running together in long recordings.
• Fixed learned corrections occasionally not applying in Korean and Japanese, or to words with different letter case.
• Fixed clipboard contents not being restored after pasting on Windows remote desktop / virtual machines.
• Fixed leftover processes occasionally remaining after quitting or restarting for an update.

• History now cleans itself up automatically: entries older than your chosen retention window (7 / 30 / 90 days / forever) are purged once a day, no manual tidying needed.

• The "Recommended for you" banner now also appears on the rewrite model tab, matching the speech model tab's experience.
• During onboarding, speech models are grouped and recommended by language, while rewrite models show as one flat list; fixed a few minor languages being grouped under the wrong neighboring language.
• Interface copy across languages is more concise and consistent, with several mistranslations corrected.
• Each settings tab now has a short intro line explaining what it covers.

• Fixed dictated text occasionally carrying over part of the previous sentence.
• Fixed polish/rewrite silently stopping after an unexpected crash, requiring an app restart to recover.
• Fixed clipboard restore occasionally failing on Windows, which could scramble the pasted result.
• Fixed push-to-talk recording sometimes stealing focus from other input fields.
• Fixed a few interface labels pointing to the wrong settings screen or not matching the actual filtered results.
• Fixed extra spacing appearing in titles on Traditional Chinese and Japanese.

• Engine details in the sidebar show the active speech and polish models and memory use.
• Added the Qwen3-ASR 1.7B speech model and Gemma 4 E2B local polish model.

• Speech recognition feels more consistent, with more coherent transcripts.
• Speech corrections rely more on your dictionary and current context; manually corrected terms can guide future recognition.
• Structured polish preserves paragraph hierarchy, making topic changes clearer.

• Contextual rewrites are steadier: decisions focus on retrieved candidates, no more wild shots.
• Auto-learning now has a safety gate: rules that conflict with your corpus are pruned in the background; manual rules go through the same gate.
• Interface text aligned with the latest design specs, including the model picker and compatibility note.

• Structured formatting now preserves perspective, tone, and uncertainty more accurately, with more natural paragraphs and lists.
• Translation now cleans up the text before translating, producing clearer bilingual output and reducing truncation in long content.

• Square brackets in Markdown and technical text are no longer removed by mistake.
• Cloud polishing no longer makes unnecessary retries on authentication, rate-limit, or service errors.

• Optimized the prompt structure of the polish system for clearer and more accurate semantic refinement.
• Added recording protection to keyboard shortcut binding to prevent accidental activation.
• UI menu items and release notes now fully support six interface languages: Chinese, English, Traditional Chinese, Japanese, Korean, and Spanish.

• Clicking the window close button now hides the app to the background instead of quitting — reopen from Dock.

• Occasional delay in text injection (results now appear faster after releasing the key).
• Output semantic reversal on the 1.5B polish model — content no longer says the opposite.
• Language detection misjudgment in rare cases.
• Model server startup failure caused by port conflicts.
• Multiple stability fixes.

• Cloud polishing occasionally leaving stray tag characters in the result.

• Built-in polish style names and descriptions now display in your interface language.
• Text insertion now uses native paste — faster and more reliable.
• Interface details refined to match the latest design.
• Removed the redundant "Finish — I'll wait" button while models download during onboarding.
• The release notes dialog now groups changes into Added / Improved / Fixed sections for easier reading.

• On some devices the app language ignored the system setting and always showed English.
• Some keys, when set as the dictation shortcut, would also trigger their original action in the active app.
• The recording capsule could occasionally jump or get clipped at the bottom edge.
• Dictating in another app incorrectly lit up the trial card on the home screen.

• Real-time voice input: hold the shortcut to speak; release and your words appear instantly at the cursor.
• Local AI polish: strips filler words ("um", "uh", "like") and splits sentences by meaning, adding punctuation.
• Smart dictionary: say a term a few times and it remembers your proper nouns.
• Domain glossaries: pick your field for more accurate recognition of technical terms.
• Multilingual: Chinese, English, Traditional Chinese, Japanese, Korean, Spanish — mixing English into Chinese won't break recognition.
• Cloud polish (optional): connect an OpenAI-compatible service when you want stronger results.
• Recording capsule: an on-screen indicator showing listening / transcribing / inserted.
• Windows support: runs on both macOS and Windows.
• Auto-update: upgrade in-app when a new version is out.