Changelog
What changed, when it shipped, and what is available now.
This page records features that have shipped and remain available in chunks.md. Dates are release dates, not announcements for unfinished work.
2026-07-20 — New OCR models and Abort
- Added PaddleOCR-VL 1.6 in INT8 and FP32 without removing PaddleOCR-VL 1.5.
- Added GLM-OCR in Q4F16, INT8, FP16, and FP32.
- Added Granite-Docling in Q4F16, Q8, Q4, FP16, and FP32 for WebGPU devices. Q4 is the recommended starting point; Q4F16 and FP16 are selectable but not browser-completion-verified. English is primary, while Japanese, Arabic, and Chinese support is experimental.
- Added an Abort control to OCR, Suno, Cutter, and test runs so active work stops and model resources are released.
- Added model guides for PaddleOCR-VL 1.6, GLM-OCR, and Granite-Docling.
2026-07-03 — Document combining and clearer OCR output
- Added the Combine documents tool for merging supported files into one AI-ready document.
- Improved raster OCR preprocessing with contrast adjustment and binarization controls.
- Improved PaddleOCR-VL decoding and row grouping so multi-column and table-like output is less likely to arrive in the wrong reading order.
2026-06-25 — Cleaner Unlimited-OCR markdown
- Added Strip layout positions to Unlimited-OCR advanced settings. It removes coordinate tokens when you want cleaner markdown instead of layout metadata.
2026-06-24 — Unlimited-OCR
- Added Unlimited-OCR, a DeepSeek-based local WebGPU option for complex document pages.
- Added a dedicated model guide covering browser requirements, first-run downloads, and practical use.
2026-06-20 — Suno instrumental modes
- Added Instrumental Light for fast center-channel vocal reduction.
- Added Instrumental Hard for stronger Demucs-based source separation when the light mode leaves too much vocal content.
2026-05-05 — Suno Audio Anonymization
- Added local audio anonymization for changing a track before reuse while keeping the workflow inside the browser.
2026-03-19 — Model precision choices
- Added Q4, FP16, and FP32 choices for SmolDocling 256M.
- Added INT8 and FP32 choices for PaddleOCR-VL 1.5 and MangaOCR.
2026-03-18 — PaddleOCR mobile
- Added PaddlePaddle-OCRv5 Mobile as the smaller, faster PaddleOCR option.
2026-03-16 — Embedded PDF text
- Added PDF Embed Text for extracting an existing PDF text layer without running OCR.
2026-03-15 — MangaOCR
- Added MangaOCR for Japanese manga and comic text.
2026-03-14 — PaddleOCR-VL 1.5
- Added PaddleOCR-VL 1.5 for structured document OCR.
2026-03-13 — First OCR models
- Added PaddlePaddle-OCRv5, PaddlePaddle-OCRv3, and SmolDocling 256M.
Related reading
What is chunks.md?
PaddleOCR v5, SmolDocling, MangaOCR, DeepSeek-based Unlimited-OCR — 8 OCR models running in your browser. PDFs and scans to markdown, zero uploads.
DeepSeek OCR online
Try DeepSeek-style OCR online: Unlimited-OCR (DeepSeek-based) runs in your browser with WebGPU. Layout-aware, multilingual, markdown out — no upload.
Suno upload fix
Suno flagged your upload as a known work? Anonymize the audio locally with this free browser tool and re-upload. No servers — your file stays on your device.