What’s new in 1.3.0 • Whisper Base is now included Base is the default for new installations and the complete model ships with the app, so you can start an offline transcription without a separate download. Bundled and downloaded model lookup is also fixed, preventing unnecessary re-download prompts. • A new recommended high-quality Turbo A compact ~632 MB Large V3 Turbo is now the recommended high-quality download. The original ~1.6 GB Large V3 Turbo remains available as a compatibility option. • More useful progress Transcription and translation now show the active stage, completed counts, model, and language details. Cancel is easier to reach, and failed or canceled work no longer leaves stale progress running. • Refined Processing Options and Settings Languages, Whisper, translation models, and download status now have a clearer hierarchy. It is easier to see what is ready, what needs a download, and what action to take next. • Updated translation models Local models add Qwen 3.6 27B and 35B-A3B. OpenAI options add GPT-5.6, 5.5, and 5.4; Gemini adds 3.7, 3.6, and 3.5; Ollama choices are refreshed too. • Project choices are restored correctly Reopening a project restores its Whisper model, translation method, and provider instead of silently using the most recent global selections. • More reliable MKV and background work Canceling MKV preparation now terminates FFmpeg and cleans temporary files. Failed or canceled transcription also stops its background progress task cleanly.

What's new in 1.2.1 • First-run onboarding On first launch you get a one-page welcome (the three-step flow plus "fully offline, no API key" — one click takes you straight to picking a video). The first time you enter the workspace, a two-step tour points at the language settings and the translation model. After your first transcription, a one-time tip on the Export button shows what's next. Everything appears once; re-watch via Help → Show Onboarding Tour. • Pre-download translation models from Settings Settings → Models now lists every built-in model: download the ones you want ahead of time (live progress, cancelable), see sizes of downloaded ones, delete what you don't need. Old pre-upgrade models are tagged and cleanable. No more waiting for a multi-GB download the first time you hit Translate. • Fixed: downloaded Whisper models showed as "not downloaded" after relaunch The model was right there on disk, but Settings offered the Download button again and the workspace warned about a pending ~500 MB download. The app was scanning the wrong directory level; this version checks where WhisperKit actually stores models (and still finds legacy locations). • Fixed: false "translation model not downloaded" warning The workspace used to demand a download whenever the model wasn't loaded in memory — even right after you downloaded it in Settings. Downloaded-but-not-loaded now shows a quiet green note (it loads automatically when translating); the warning only appears when the model is truly missing. • Whisper download heads-up Picking a Whisper model that isn't on disk now tells you up front ("about X will be downloaded when transcription starts"), and the progress message honestly says downloading vs loading. • Default translation model is now Qwen 3.5 0.8B Lighter and faster than 2B, and good enough for most subtitle work — the "Recommended" tag moved with it. 2B is still one click away when you want more quality. • Test Connection for Gemini The Gemini API key now has a Test Connection button, same as OpenAI. The API tab footer also clarifies that keys are only needed for cloud translation — the built-in offline models need none. • Fixed: About always said "Version 1.0" It now shows the real version. • Window opens at the right size The main window now opens at workspace size instead of starting narrow and jumping wider once a video loads.

What's new in 1.2.0 • New: Gemini cloud translation One more translation engine — Google Gemini (gemini-3.5-flash, gemini-3.1-pro). Paste your Gemini API key under Settings → API to use it. • Translation model picker, unified You used to pick a "method" first, then a model. Now there's a single "Translation Model" dropdown grouped by type (built-in offline / Gemini / OpenAI / Ollama / custom API) — the engine is inferred from whatever model you pick. Cloud models that need an API key are marked with a key icon. • Settings, simplified Settings used to show every method's model field at once, most of them irrelevant. Model choice now lives in the workspace and remembers your last pick; new projects reuse it. • Built-in local models refreshed - Qwen 3 across the board → Qwen 3.5: 0.8B, 2B (default), 4B, 9B - Gemma 3 4B → Gemma 4 E4B, added Gemma 4 E2B, kept Gemma 3 1B - Phi-4 mini unchanged Note: the model IDs all changed, so picking a new model downloads it once. • Fixed: blank video when reopening a project from history On first open the app wasn't saving a sandbox access token (a security-scoped bookmark), so on reopen macOS refused to let it read the file and the player went black. Subtitles were fine because they live in the project file. This version creates and saves that token on first open. (For projects already in your history, open the video once via "Open Video…" to heal them.) • Export / Burn now ask where to save when you press the button You used to pre-select an output folder in options. Now Export and Burn open a save panel (defaulting to the video's folder, with the filename filled in). This also fixes export/burn failing after relaunch because the saved folder permission was gone. • Video load failures are now visible If a video can't be read, you get a clear message instead of a silent black screen. • Subtitle editing, easier Click a line's text to edit it inline — the whole line is the hit target. Click the timestamp to edit timing in a popover (with a "jump to time" button). The number/▶ on the left jumps the video to that line. • Fewer skipped translations Models occasionally drop a few lines, leaving them stuck on the original text. This version auto-retries the skipped ones (smaller batches, then one-by-one) and matches results more robustly (case-insensitive ids, positional fallback). • Clean up local models in Settings Settings → Models now lists your downloaded offline models with their sizes, with a one-tap "remove unused" for last release's models (Qwen 3 / Gemma 3) to reclaim disk. Deleted models re-download automatically when next used.

What's new in 1.1.3 • Fix: local translation models re-downloading on every launch MLX was caching downloaded models in ~/Library/Caches, which macOS happily purges whenever disk space gets tight — so the next time you opened SRTWhisper, the model had to download again. Moved to Application Support, where the OS leaves it alone. Anything you already downloaded migrates automatically on first launch (one-time, won't run again). • Local translation model lineup, refreshed - Qwen 2.5 family replaced with Qwen 3: 1.7B, 4B (default), 8B - Added Gemma 3 1B and 4B (QAT-quantized, multilingual, especially strong on European languages) - Added Phi-4 mini (English-focused, better at idioms and reasoning) Picker now has 6 options instead of 3. Default upgraded from Qwen 2.5 1.5B to Qwen 3 4B — noticeably better multilingual quality. • Bundled Whisper model bumped to base The bundled model used to be tiny — fine for English, rough on Chinese and Japanese. 1.1.3 ships with base (about 80 MB heavier app), so the first transcription you run produces a reasonable result without waiting on a download. • OpenAI model picker, current with 2025 Added GPT-5, GPT-5 mini, GPT-5 nano, and o3 (full). Default upgraded from gpt-4o-mini to gpt-5-mini. • Ollama default updated to qwen3:4b Matches the local LLM side, so switching between MLX and Ollama gives you similar translation quality. The old default (llama3.2) was two generations behind. • Detail - Translation model loading UI now reads "Downloading" (first time) versus "Loading" (subsequent launches), instead of always saying "Downloading"

What's new in 1.1.2 • MKV support, bundled You used to have to convert MKV files in Handbrake first. Not anymore — just drop them in. SRTWhisper bundles an LGPL build of FFmpeg that remuxes the container to MOV on the fly. In most cases the video and audio streams are copied without re-encoding (takes a few seconds). If the audio is in a codec QuickTime can't decode directly (Vorbis, Opus, etc.), the audio is auto-converted to AAC — video still copied, no quality loss. You can turn this off in Settings → MKV Support if you don't need it. • Subtitle style preset gallery There's a new "Style" button in the toolbar. Click it and you see a row of preset style thumbnails — pick one and it applies uniformly to every subtitle. No more one-by-one tweaking of font, color, and outline width. Switch to Customize mode if you want fine control. • English & Japanese localization, finished properly Some alerts, progress messages, and quality reports were still showing Chinese in the English and Japanese builds. This release sweeps every user-facing string, so the translations are complete now. • Other - Better Intel Mac transcription stability warning - Quality check report improvements

Stability notice for Intel Macs. If you're running SRTWhisper on an Intel Mac, a banner now warns that transcription may be unstable. This is due to a known limitation in the underlying speech recognition library on Intel hardware. Apple Silicon are unaffected and remain the recommended platform.

Improved localization across Traditional Chinese, English, and Japanese. Refined progress display during model loading, toolbar layout, and translation error messages.