NextAI Translator: vanhemmat versiot ja versiohistoria

v0.6.55
- Windows OCR: recognize Chinese (and other installed OCR languages) when English is the first Windows display language; previously only digits and Latin words came back, or nothing (#1902, #1950, thanks @zhouzhendiao)
- Windows OCR: no more spaces between every Chinese/Japanese character in recognized text (#1950)

v0.6.54
- Fix the API Model dropdown opening invisibly (zero width) after switching provider in settings; affected every provider's model selector, reported for Gemini (#1945, #1947, thanks @Loongphy)
- Gemini: send the thinking config each model accepts (uses the models API's thinking flag, with per-family fallbacks), fixing HTTP 400s on several models (#1947)

v0.6.53
- Gemini: the API Model selector now shows why the model list could not be loaded (invalid key, region, quota...) instead of an empty dropdown, and only lists models that can translate (#1945)

No more Japanese → Japanese "translations"
- When the detected source language is the same as your default target language (for example Japanese text with Japanese as the default target), the translator now picks a different target instead of translating into the same language: your writing target language if it differs, otherwise English. Applies to the translator window and the browser extension's inline lookup (#1873, thanks @hrtshu)

Gemini follows the thinking switch, and MiniMax-M3 is the default MiniMax model
- Gemini now follows the "thinking" switch in settings: with it off, translations use as little thinking as each model allows (faster); with it on, the model thinks normally. Gemini 3.x models use their thinking levels (#1900, thanks @Loongphy)
- MiniMax-M3 is now the default MiniMax model; the older M2.5 models were removed from the list (your saved model choice is kept) (#1884, thanks @octo-patch)

The end of translations no longer goes missing with OpenRouter and some other providers
- With some providers, notably Gemini via OpenRouter, the last sentence or paragraph of a translation could be cut off. It is now always shown (#1810, #1923, thanks @Hyhyhyyy)
- Cerebras now shares the common OpenAI-compatible streaming code, so it gets the same fix (#1923)

macOS: typing works again in the translator window
- On macOS, after the translator window was shown again (for example via the hotkey), the input box could stop accepting keyboard input until you clicked it. Keyboard focus now returns to the window automatically (#1915, #1944)

Current Claude and Azure models work again, and failed lookups stop spinning
- Fixed Claude Opus 4.7+, Sonnet 5+, Haiku 5.5 and Fable 5+ rejecting every request: they no longer accept temperature or a thinking budget, so the app now uses adaptive thinking with an effort level on those models, and the 5.5 models are in the model list (#1937, #1941)
- Fixed Azure OpenAI rejecting gpt-4o, gpt-4.1, gpt-5 and o-series deployments, and Azure AI Foundry endpoints, by sending max_completion_tokens instead of max_tokens (#1926, #1936)
- Fixed the Responses API receiving max_completion_tokens; it is now sent as max_output_tokens (#1942)
- The compact inline lookup window now stops loading and shows the error when a translation fails, instead of spinning forever (#1938, #1942)

Shorter wait for the first character, and no more garbled Chinese
- Shortened time to first character: the HTTP client is now reused between translations instead of paying a fresh DNS + TCP + TLS handshake each time, the first streamed token is painted immediately instead of waiting out the 120ms batching window, and settings are cached instead of re-read from disk on every request
- Fixed streamed translations that could start mid-sentence because the first chunks arrived before the desktop app had registered its listeners
- Fixed multi-byte characters turning into replacement characters when a network chunk split them, which affected most Chinese, Japanese and Korean translations
- Fixed the desktop abort listener leaking when a request failed, and a few panics on malformed request options
- Added time-to-first-token tracing: each translation reports where its time went - prompt, settings, network setup, provider thinking, first paint - in the devtools console, in the app's stdout, and through window.__translatePerf

Fix newer GPT-5 models being sent an unsupported parameter
- Fixed newer GPT-5 models (such as gpt-5.0) being sent a reasoning effort setting they no longer support, which could make requests fail (#1853)
- Refreshed the built-in fallback model list, shown when the model list cannot be fetched from the API, to lead with the GPT-5 series instead of GPT-4 models

Fix the model picker committing the wrong model
- Fixed the model combobox sometimes selecting and saving a different model than the one clicked in the filtered dropdown list

Reliability fixes for submitting and loading states
- Fixed Enter being swallowed while using a Chinese/Japanese/Korean input method: the keystroke that commits the IME candidate no longer requires a second Enter to submit (#1919)
- Fixed the UI getting stuck on "Translating..." forever after switching languages or clearing OCR text mid-translation (#1919)
- Submitting now falls back to the first available action and shows a clear message when the built-in translate action was renamed or deleted, instead of doing nothing (#1919)
- Fixed background requests in the browser extension hanging forever on connection errors (#1918)

Browser-like back and forward navigation between translation views
- Added browser-like back/forward navigation: clicking a word in the translation result to view its details can now be undone instantly with the back button in the bottom bar, Cmd/Ctrl+[ and ] shortcuts, or the mouse side buttons
- Going back or forward restores the previous view's result instantly from an in-memory cache, without re-running the translation

LiteLLM provider, Markdown tables, and a Linux tray fix
- Added LiteLLM as a built-in AI gateway provider
- Markdown tables (GFM) in translation results now render as real tables instead of raw text
- Fixed the Quit action missing from the tray menu on Linux
- Fixed the Tauri Vite build on Windows

Reasoning model support in the model list and completed Korean translations
- The OpenAI model list now includes o-series reasoning models such as o1, o3 and o4-mini
- Completed the missing Korean translations for the dictionary feature

Adjustable UI font size and a clear error message when WebView2 is broken
- Added a UI font size setting so the whole interface can be scaled to your preference
- On Windows, if the Microsoft Edge WebView2 Runtime is missing or broken, the app now shows a clear error dialog with a button to open the official repair download page, instead of silently flashing a window and exiting

Fix the app permanently failing to launch after a settings file corruption
- Fixed the app never opening again (window flashing and closing instantly) once its settings file got corrupted, e.g. by a crash or power loss during a save: a broken settings file is now backed up and reset at startup instead of crashing the app before any window appears
- Settings are now saved atomically (write to a temporary file, then rename), so an interrupted save can no longer corrupt the settings file in the first place

Major stability release: fixes freezes and progressive slowdown
- Fixed the app becoming slower and slower the longer it ran, eventually freezing: an error-tracking session recorder leaked resources under this app's heavy UI updates and has been removed
- Fixed random unresponsiveness where Submit and other actions did nothing for seconds or minutes: an internal blocking sleep could starve all app commands
- Fixed streamed translations leaking event listeners every time a translation was superseded
- Summoning the translator window now always brings it to the front, even over other apps' windows
- Fixed windows staying throttled and unresponsive after being covered and revealed again
- The API model selector is now a combo box: type any model name freely or pick from the filtered list
- Added OpenRouter as a built-in provider with dynamic model listing

Major responsiveness fix for streaming translations
- Fixed the UI freezing and buttons not responding while a long translation is streaming in
- Long example sentences in dictionary results now wrap instead of forcing the result area to scroll sideways
- Reading highlight timing for local TTS is noticeably tighter: syllable-based estimates, punctuation pauses accounted for, and highlights lead the voice slightly instead of trailing it

Fixes the Submit button sometimes doing nothing
- Fixed the Submit button silently doing nothing right after restoring a history entry that was already loaded
- Pressing Enter to resubmit unchanged text now re-runs the translation instead of doing nothing

OCR works on macOS again
- OCR on macOS now uses the built-in screenshot selection with native Apple Vision recognition, replacing a broken helper binary that failed silently
- Fixed OCR randomly doing nothing on all platforms: the capture window suspended its own logic before recognition could run
- Retrying OCR after a failed attempt now works instead of silently doing nothing
- macOS will ask for Screen Recording permission on first use

Performance fixes for UI lag
- Fixed OCR workers leaking after rejected image drops, which gradually made the whole UI laggy the longer the app ran
- Typing in the input box no longer recomputes the highlight overlay on every keystroke when nothing needs highlighting
- Fixed a leaked audio pipeline after failed speech synthesis that kept the app burning CPU in the background
- The OCR spinner no longer gets stuck when recognition fails

Fixes a gray artifact at the window corner
- Fixed a gray artifact that could appear sticking out of the window's top-left corner when background blur is enabled for a long-running app
- The native blur layer no longer stacks up on theme changes and now follows the window's rounded corners

Improvements and fixes for the reading experience
- Reading highlight now stays in sync with the voice: Edge TTS uses the
service's exact per-word timings, and local TTS timing accounts for
leading/trailing silence
- Word definition hover cards now work inside example sentences, and the
hover card and word highlight have a cleaner, more refined look
- Numbered example sentences without punctuation after the number are
now recognized as readable
- Reading, hover, and input-box highlights share one accent color
palette with smooth transitions
- Fixed the loading spinner wobbling around an off-center origin
- Fixed the language swap icon vertical alignment
- Redesigned the footer update pill
This release also includes the word definition hover cards (#1899),
which were previously only on main and never shipped in a release.

v0.6.27
- feat: show a neutral footer update button when a new version is available
- feat: open the updater directly from the translator window

v0.6.25
- fix: vertically align example speech controls

v0.6.22
- feat: add TeamoRouter as a built-in LLM provider
- i18n: add Korean (ko) translation

v0.6.21
- Fix model-list fetching and translation failing on the desktop app when the API URL includes an explicit port (#1571). Every desktop request goes through the Tauri HTTP allow-scope, which used http://* / https://* patterns. URLPattern only matches those against default-port URLs, so any host with an explicit port (e.g. a local server at http://192.168.x.x:11434, or a host like https://example.com:1234) was rejected before the request was ever sent. The scope now uses http://*:* / https://*:*, matching URLs both with and without a port.

v0.6.20
- Actually fix the high idle CPU usage on Windows that persisted through 0.6.19 (#1886). Hiding a window in Tauri only hides the OS window — it never tells the WebView2 controller it is hidden, so the renderer of the always-alive Quick Translator and writing-indicator panels kept compositing at full frame rate even while the app sat idle in the tray. The WebView2 controller is now explicitly suspended (SetIsVisible) on every hide, and the two panels are no longer pre-created at startup on Windows.