Fixed an issue where some Bilibili videos could not start processing after the bundled downloader was updated.
v1.0-rc16
Sumari's Dock icon now appears at the same visual size as macOS system icons instead of looking oversized.
A summary service can now save multiple models. Defaults, retries, and recovery choices select the exact service and model instead of treating the whole service as one option.
Each configured summary model now has its own connection test, context window, output limit, and reasoning settings.
Dragged files, pasted links, browser imports, playlists, and other batches now share one ordered queue. Users can choose 1–3 simultaneous videos, and waiting for a service choice no longer blocks unrelated videos.
Queued and running videos can be canceled. Sumari waits for background work to actually stop before reusing the slot, preventing hidden overlap.
Dragged and pasted items keep their accepted order even when macOS finishes reading individual files out of order.
If a saved service configuration is damaged or comes from an unsupported future version, Sumari now preserves the original data, temporarily locks related settings, and offers export or explicit reset instead of silently overwriting it.
Incomplete, refused, paused, tool-directed, or truncated AI responses are no longer saved as finished summaries. Partial text is discarded while the completed transcript remains available for retry.
Service and database safety is tighter: API credentials cannot follow redirects to a different server, raw service failures no longer leak into history or diagnostic logs, and any database upgrade that removes duplicate records requires a verified backup first.
Compatibility note: older Sumari versions read only the first model saved for each summary service. Opening and saving that service in an older version removes its additional models.
v1.0-rc15
Sumari now checks for updates once when the app launches, as long as automatic update checks are enabled, for both stable and development-channel updates
Custom OpenAI-compatible summary models now receive only the fields required for each request mode, fixing strict models such as Kimi K3 that reject a hard-coded temperature or other unsupported optional fields
Streaming summary recovery now stays within the configured retry limit and keeps the provider, model, and streaming choice frozen for that task instead of retrying repeatedly or changing configuration mid-task
Built-in cloud presets now use current recommended models, matching regional service addresses, and verified context windows, including DeepSeek V4, Kimi K3, MiniMax M3, GPT-5.6, GLM-5.2, Qwen 3.7, Groq GPT-OSS, and Gemini 3.6
ElevenLabs Scribe and other strict speech-to-text services now receive provider-compatible upload fields for both direct and segmented transcription, with one shared retry limit across the whole request
When Chrome has multiple profiles, Sumari now asks you to choose the correct Download Authentication profile instead of guessing which signed-in account can play the video
v1.0-rc14
Fixed drag-and-drop breaking entirely: dragging links onto the window now keeps working instead of stopping after a few drops and requiring an app restart to recover
The paste-link field now adds the link when you press Return, not only when you click the button on the right
Added a "Share anonymous usage statistics" toggle under General → Privacy (on by default) to help improve Sumari; it reports only limited feature-usage signals and never video links, titles, transcripts, or summaries, and can be turned off at any time
First-run onboarding now notes that anonymous usage statistics are on by default and points to General → Privacy to turn them off
Hardened connection safety for API-key providers that use local or self-hosted endpoints; a few setups on insecure local addresses may need to switch to HTTPS or a specific IP address
The transcription-language hint moved to transcription settings (previously under summary settings), and its copy clarifies that it only affects local Apple Speech / Whisper audio transcription, not the summary language
Fixed formula-like text inside code blocks in summaries possibly being rewritten incorrectly
A few security and stability improvements
v1.0-rc13
The TLDR at the top of each summary is easier to skim: tutorials and tip videos use a numbered, timestamped checklist, while narrative, interview, and analysis videos use short paragraphs instead of one dense block
Long summaries now collapse individual detail chapters by default while keeping the TLDR visible; custom prompts and malformed summaries still use full-text rendering
Summary cards use a stable background, so very long streaming summaries no longer flash for a frame
You can drag an external link or local media file straight onto the summary body to start an import
The main window behaves more like Notes: a sensible sidebar width range, system collapse/restore behavior, no unexpected widening while dragging the divider, and no oversized top gap when hidden
The sidebar search field and link input no longer flash dark while the sidebar shows or hides
Fixed ElevenLabs transcription running words together for English and other space-separated languages (which also hurt summary quality; Chinese, Japanese, and other providers were unaffected)
"Retry from scratch" no longer re-downloads and re-transcribes the video when it then asks you to choose a summary provider
After the last usable summary provider is removed, the selector no longer shows a blank state and falls back to "ask every time"
Transcription failure messages no longer include raw technical details from the provider
If the local database cannot be opened after an update, the app stays usable and offers a safe "back up and reset" recovery instead of failing to launch
Bilibili links are more reliable, including public videos the site previously rejected before Sumari could read their metadata
A few security and stability improvements
v1.0-rc12
Export now opens a preview sheet: review Markdown or PDF first, then choose thumbnail, description, video info, chapters, and transcript with checkboxes
PDF export adds font, body-size, and margin controls, with Chinese first-line indent and more book-like heading spacing
The default summary template now focuses on a fuller TLDR, detailed content, and terminology; the old brief/detailed split is gone
The reader and PDF export no longer mistake dollar amounts for formulas, and long summaries scroll more reliably
Sidebar search now covers errors, transcripts, video descriptions, channels, tags, and chapter titles
The API setup guide has fresher defaults for OpenAI, DeepSeek, SiliconFlow, Kimi, and other providers
Apple now shows a green “Transcription only” state when Apple Speech works but Apple Intelligence is unavailable
The add-provider preset grid is now a simpler dropdown picker
Error messages and diagnostic reports are more careful about hiding API keys, local paths, and other sensitive details
A few smaller fixes and stability improvements
v1.0-rc11
Bilibili extraction is more reliable, and X/Twitter posts without playable video now fail clearly
YouTube cookie and bot-check retry handling is more resilient to transient network problems
Mixed batches keep using platform subtitles first even when some items fall back to audio transcription
Retry and recovery sheets preserve the active task owner more reliably, reducing duplicate processing and stuck interrupted states
Sidebar thumbnails refresh actively, so the list no longer needs manual scrolling before images appear
Share and export checks catch missing fonts or assets earlier
English localization covers more settings, export, admin, provider, and recovery paths
Bundled tools and release packages are more consistent, including QuickJS handling and local preparation for yt-dlp, ffmpeg, quickjs, and whisper.cpp
v1.0-rc10
Summaries now skip sponsor and ad segments more reliably without leaking instruction-like text into the output
Long summaries have more stable heading structure with fewer broken Markdown heading artifacts
Videos without subtitles still fall back to transcription, but real subtitle tool or network failures now surface directly instead of being silently masked
PDF export is more reliable for long transcript-heavy entries
Networking is simpler: Sumari now follows the macOS system proxy instead of keeping a separate in-app proxy configuration flow
v1.0-rc9
YouTube bot-check failures now retry automatically and can refresh browser cookies before restarting the flow from the beginning
Fixed stale cookie-cache behavior and now show the real cookie refresh time in Settings
Processing-time network requests now consistently follow the frozen configuration used for the active run
Also fixed Ollama readiness checks, bundled-tool update scripts, database cleanup, and missing English copy
v1.0-rc8
Summary and PDF rendering language now follows the interface language consistently
Chinese serif font fallback is more reliable across the reading view and PDF export
Browser capture is more forgiving — if one browser denies automation, others keep working
Sidebar thumbnails load faster with proper caching on repeated views
v1.0-rc7
Summary rendering follows the interface language consistently for more predictable mixed-language reading
Chinese typography fallback is more consistent in both the reading view and PDF export
Browser capture keeps working even if one browser refuses automation access
Bundled font handling and release packaging are more reliable
v1.0-rc6
Clearer error messages when subtitle fetching, audio download, or provider setup fails
More reliable restart after changing the interface language
Safer quit behavior when work is in progress, with less chance of losing state
Whisper model downloads are more reliable and properly clean up on failure or cancel
Some AI providers now produce cleaner summary output with fewer formatting artifacts
v1.0-rc5
Added per-site browser capture controls — disable auto-capture for specific sites
Added "Force Analyze" for unsupported links to try yt-dlp anyway
Improved long-form reading in the detail view
Refined serif summary typography for more comfortable reading
v1.0-rc4
Simplified retries into clear "start over" and "regenerate summary" actions
Unified provider selection into clearer transcription and summary flows
Improved Whisper model download and recovery in the selection sheet
More predictable retry behavior when subtitles are available or not
v1.0-rc3
Fixed a startup crash on macOS 14 (Sonoma)
Kept Apple Intelligence available on newer systems without breaking older supported Macs
v1.0-rc2
Bundled QuickJS so YouTube extraction works more reliably out of the box
Whisper offline mode is easier to discover before models are downloaded
v1.0-rc1
First public test release: full browser capture, link pasting, playlists, history, detail view, and retries
Multiple transcription and summary paths: local Whisper, Apple Speech, and cloud providers
Markdown and PDF export
Bilingual interface: Simplified Chinese and English