v0.10.0

Breaking changes

• remove the Document tab — the app translates text only, and liteparse goes with it

Features

• design the Reading Room theme — warm dark-first palette, AA button label, bundled fonts
• audit fixes and design pass across both tabs
• switch to Streamlit's default theme in light and dark

Fixes

• keep the typed draft and the character cap across a swap, and give the output state two writers
• drop the subtitle its own model load contradicted, and keep the Hub token and telemetry out of that request
• silence load_model's second spinner, and correct the record of how long the phantom gap lived
• stop both tabs shifting 16px on Translate — give the transient spinner an st.empty() slot
• make "nothing is sent to a server" true for a page load, and harden the theme guards
• label document read failures honestly and name a scanned PDF's missing text layer
• report a failed image OCR instead of "No translatable text found"
• accept .tif uploads alongside .tiff
• resolve code-review findings from the audit and design pass
• keep the output panel one shade through streaming
• report tokenizer failures in the warning slot
• show the loading spinner where the user is looking

Performance

• render the UI before loading the model

Documentation

• state both input limits without predicting the script, and put the tab-label record in the past tense
• test the main display for the hero and record why it still cannot frame it
• retire the Document tab from CLAUDE.md and the theme's comments
• regenerate the README hero without the tab strip
• regenerate the README hero without the subtitle it was still advertising
• promise that nothing you translate leaves your Mac, and say what the app does fetch
• note the watcher tracebacks TRANSFORMERS_VERBOSITY does not cover
• re-verify CLAUDE.md against Streamlit 1.63 and liteparse 2.14
• group the CLAUDE.md conventions and add the single-test commands
• correct the ty-ignore scope and complete the structure map
• record how to regenerate the README screenshot

Full changelog: v0.9.0...v0.10.0

v0.9.0

Highlights

• 📄 Document parsing rebuilt on LiteParse. Docling is gone — parsing is now one ~11 MB wheel with zero transitive dependencies, replacing ~278 packages and PyTorch.
• 🔒 PDFs never touch the network. PDFium and Tesseract are statically linked, so nothing needs to be on your PATH, and born-digital PDFs parse from their text layer in ~0.04 s with zero network calls.
• 🖼️ Image translation with source-language OCR. PNG/JPG/TIFF/WebP are OCR'd in the language you pick as the source rather than always English. Each OCR language fetches ~12–15 MB of Tesseract data once, then works offline.
• ✂️ Token-aware document chunking. Chunks carry their enclosing markdown headings as context, split on Latin and CJK sentence boundaries, and never cut a multi-byte character in half.
• 🌍 67 languages, text and documents, streaming output, entirely on your Mac — no API key, nothing leaves the machine.

⚠️ Upgrade notes

• The docs extra is gone. uv sync --extra docs now fails — run plain uv sync. Document support is built in.
• Office and HTML uploads are no longer supported. v0.8.2 accepted DOCX/PPTX/XLSX/HTML through Docling. LiteParse needs LibreOffice on PATH for Office formats and does not handle HTML, so uploads are now PDF and images only: pdf, png, jpg, jpeg, tiff, webp. Convert other formats to PDF first.

Requirements

Apple Silicon Mac · 8 GB+ RAM · Python 3.13+ · uv

Run

uv sync
uv run streamlit run streamlit_app.py

First run downloads the model (~3.6 GB). Loads mlx-community/tiny-aya-global-8bit-mlx — CC-BY-NC, non-commercial only.

---

Features

• add LiteParse as a second document parser backend

Fixes

• bound the heading trail and ignore headings inside code fences
• preserve multi-byte characters and OCR in the source language
• repair document chunking and image parsing

Refactoring

• drop Docling, make LiteParse the only parser

Documentation

• add screenshot and badges
• lead with value, fix test command, surface non-commercial license
• document release workflow in CLAUDE.md

Full changelog: v0.8.2...v0.9.0

v0.8.2

Highlights

• 🪪 Application code is now licensed under Apache-2.0 (the tiny-aya-global model weights remain CC-BY-NC, non-commercial only).
• 🌍 Translate text and documents across 67 languages with streaming output.
• 🍎 Local inference on Apple Silicon via MLX — no API key, nothing leaves your machine.
• 📄 Document translation (PDF, DOCX, PPTX, XLSX, HTML) via Docling — optional docs extra.
• 🎨 Nord theme with light and dark modes.

Requirements

Apple Silicon Mac · 8 GB+ RAM · Python 3.13+ · uv

Run

uv sync                # core
uv sync --extra docs   # + document translation
uv run streamlit run streamlit_app.py

Loads mlx-community/tiny-aya-global-8bit-mlx (CC-BY-NC, non-commercial only).