Bring your own books
Open local files. Reading position, bookmarks, notes and review history stay tied to the right title.
Built with GPT-6 Astra · WhaleRead 1.18
Read across languages.
Keep the book.
Translate local TXT, Markdown and EPUB files, read source and translation side by side, fix awkward passages, and ask AI from the page—without moving your whole library into a consumer cloud.
中文 / English interface
18 common language directions
7B on-device
Hy-MT2 adapted
WhaleRead · bilingual reading
Actual WhaleRead 1.18 walkthrough
Only the reader.
No staged book.
This 65-second recording shows the working macOS app with original synthetic passages: on-device 7B and private 30B translation, bilingual reading, live interface switching, evidence-first review, and Ask AI.
- Every passage was written for this demo
- No copyrighted book or personal library appears
- No desktop, notification, account, API key, or private endpoint is shown
Actual app recording · original test corpus · real model outputs may contain errors
Keep translation local
Run book translation on-device with Hy-MT2 7B. Ask AI uses only a compatible endpoint you configure.
Keep the last word
Model suggestions stay drafts. Nothing rewrites the book until you inspect and apply it.
A reader, not a conversion queue
Translation ends where reading begins.
Switch between source, translation and bilingual modes while safely saved work remains readable. Search, bookmarks, notes, typography and chapter position stay inside the same calm workspace.
- Read finished passages before the whole task is done
- Preserve EPUB chapters, images and in-book styling
- Resume translation from verified checkpoints
- Change the interface language without changing the book
Same book. Two private model sizes.
Fast on the Mac. More nuance on self-hosted 30B.
We ran the same five original test passages through both model paths. On-device 7B is useful and private. Self-hosted 30B more often preserves idioms, register and narrative rhythm, with quality that feels close to cloud-scale models while staying on infrastructure you control.
Qualitative synthetic-text demo, not a benchmark. Outputs below are copied verbatim.
Current download: the public preview exposes on-device Hy-MT2 7B for book translation. The 30B results below come from our private evaluation setup and are shown as quality evidence; the 30B route is not included in this download.
Based on our testing, Hy-MT2 currently provides the best fit for WhaleRead's long-form translation workflow and is the only model family fully adapted and validated in this preview. We plan to expand model support based on user feedback.
English source
“Mara did not buy it. She kept her cards close to her chest…”
The phrase means she kept her intentions secret.
On-device7B
“玛拉并不信服。她将手中的卡片紧紧抱在胸前……”
Literal idiom: readable grammar, wrong meaning.Self-hosted private30B
“玛拉并不买账。她始终守口如瓶……”
Meaning and tone survive naturally.Regional voice, not dictionary English
- estaba en las nubes
- was daydreaming
- no tenía pelos en la lengua
- wasn’t one to beat around the bush
- nunca llegó a emitirse
- never made it to air
Meaning travels, the cat stays home
- 猫の手も借りたいほど忙しい
- needed every hand available
- その話は水に流そう
- Let’s just forget about that
- 肩の力を抜いた
- relaxed
Natural phrasing in the target language
- under the weather
- 体調を崩していた
- call it a day
- 今日は終わりにしよう
- pulling their leg
- 冗談を言っているだけ
Idioms adapt to local expression
- let the cat out of the bag
- laissa échapper la vérité
- make a mountain out of a molehill
- ne pas en faire tout un plat
Review that never silently overwrites
Spot the awkward line. Let a model explain it. You decide the fix.
Highlight a suspicious translation and keep the source evidence, current wording, model reasoning and revision together. The review in this screenshot caught the literal “cards” problem. Its first suggestion was still imperfect, so the final saved revision uses the more natural 30B rendering: 始终守口如瓶.
- AnnotateSelect a phrase and describe what feels wrong.
- ReviewCompare the model’s reason with exact source evidence.
- EditKeep, rewrite or discard the suggestion.
- ApplyUpdate the reading edition only after confirmation.
WhaleRead treats model fallibility as part of the product design. Even a 30B reviewer can suggest an incomplete fix; the workflow keeps that visible instead of hiding it.
Built for people who do not think in language codes
English, 日本語, Français, Español.
WhaleRead’s interface switches between English and Chinese. Translation choices use each language’s own name, while the saved task keeps a stable identity underneath.
Ask AI without losing the page
Questions start in the passage.
Ask about an idiom, cultural reference or character. See the context before it is sent, choose the answer language, and keep useful answers with your reading notes. Answers never edit the book by themselves.
A privacy boundary you can explain
Your library stays yours.
Core reading and editing work remains on the Mac. AI requests go only to the model destination you choose.
On-device 7B
Translation runs on this Mac. Book text does not leave the device.
User-configured Ask AI
Questions go only to the compatible endpoint and credential you save locally.
Visible context
Ask AI shows the passage prepared for the request before you send it.
Human confirmation
Suggestions remain separate from book files until you explicitly apply them.
What the test set also caught
Good demos include the misses.
The self-hosted 30B handled the French idioms well, but titled a female cartographer Le cartographe silencieux while the body used la cartographe. That inconsistency is exactly why WhaleRead includes side-by-side evidence and review instead of a “perfect translation” badge.
GPT-6 Astra Challenge
Built with careful help from three AI collaborators.
GPT-6 Astra reviewed the architecture and independently audited the release. GPT-5.6 Sol kept the long multilingual implementation coordinated and coherent. DeepSeek turned tightly scoped construction tickets into working features and reran the evidence gates. Runtime translation remains user-selected: on-device 7B or self-hosted 30B.
Public preview for macOS
Language should not stand between you and a book.
WhaleRead 1.18.1 is available for Apple silicon Macs running macOS 13 or later. The 7B model is downloaded separately. The preview is ad-hoc signed and not yet notarised by Apple; the release page includes first-open instructions and a SHA-256 checksum.