emebala Ebook Reader embeds local LLM for offline translation on 3-4GB VRAM
A 1.8B translation model running locally on llama.cpp powers real-time book translation without cloud costs.
A developer frustrated by the lack of Korean translations for foreign books has created 'emebala', an ebook reader that embeds a local LLM for translation. Built on llama.cpp, it uses a specialized 1.8B parameter translation model that consumes just 3–4GB of VRAM, making it practical for consumer GPUs. The tool lets users read foreign-language ebooks and get real-time translations directly within the reader, eliminating the need for cloud APIs or external dictionaries.
The project also includes reader-centric features: sticky notes for marginalia, bookmarks with multiple tags, a book review system, and a searchable repository of all notes and highlights. The developer’s previous experience fine-tuning a 4B model and renting GPUs informed the design, aiming for a balance between quality and local performance. While the current model handles German-to-English translation convincingly, the ultimate goal is to enable translation for underserved language pairs—like Korean to English—for which no official translations exist.
- Uses a 1.8B translation model running locally via llama.cpp, requiring only 3–4GB VRAM
- Includes sticky notes, multi-tag bookmarks, book reviews, and full-text search across notes
- Designed for book lovers needing translations of non-English books (e.g., Korean, German) without cloud dependency
Why It Matters
Democratizes AI translation for niche literature, enabling offline reading of untranslated books on modest hardware.