Pluggable engines
PP-OCRv6, MiniCPM-OCR, Xiaomi-OCR-0 and PaddleOCR-VL share one task, preview and export pipeline. Capability flags absorb the differences.
Glyphix folds engine selection, model installation, batch recognition and export into one workflow. All inference runs on your machine — no uploads, and it keeps working offline.
Free · Local inference · Your files never leave the device
From downloading a model to exporting a spreadsheet — one window, no detours.
PP-OCRv6, MiniCPM-OCR, Xiaomi-OCR-0 and PaddleOCR-VL share one task, preview and export pipeline. Capability flags absorb the differences.
A JSON catalog drives the model library, with SHA-256 verification, resumable transfers and mirror fallback. Proxy changes apply to every new download.
App defaults, engine defaults, batch options and per-task overrides merge in a fixed order. Cancel a batch, retry a single item, watch every stage.
Markdown, HTML, page overlays, plain text, blocks and the raw IR stay linked. Low-confidence and uncertain blocks are visible everywhere.
TXT, JSON, CSV and Markdown. JSON keeps the full structure and Markdown matches the preview exactly.
Tray, system proxy, keychain, file associations and single instance. Themes follow the OS and the UI language switches without a restart.
Pick by scenario — lightweight text capture, complex layouts or end-to-end document parsing. Switching engines never means switching tools.
Detection, recognition and charset ship as three files. Ideal for instant text capture and low-end hardware.
A 0.75B vision-language model for curved pages, photos, tables and formulas.
0.8B covering layout analysis, key information extraction and document Q&A.
0.9B that parses tables, formulas, charts and seals straight into structured text.
Models download on demand. Verification, resuming and mirror fallback are handled by the app; shared runtimes install once.
No command line, and no shuffling files between tools.
Choose an engine and quantization in the engine center. The app downloads, verifies and swaps versions atomically.
Drop PDF, PNG, JPEG or TIFF files. The queue de-duplicates and orders them automatically.
Set a batch engine and language, override per task, cancel anytime and retry only what failed.
Verify results across six linked views, then export to TXT, Markdown, CSV or JSON.
Installers ship only the runtime. Models are installed on demand inside the app, so your disk is never filled up front.
This is the 1.0 preview. Get installers and checksums from the release page.
The questions that come up most often about on-device OCR.
No. Every engine runs inference locally. The app only goes online when you install a model or refresh the catalog.
No. Engines, models and quantization levels are installed independently, and removing one never breaks a shared runtime used by others.
The interface ships in Simplified Chinese, English and Japanese. Recognition depends on the engine: PP-OCRv6 covers Chinese, English and Japanese, PaddleOCR-VL adds Korean.
TXT, JSON, CSV and Markdown. JSON keeps the full structure, Markdown matches the preview, and pages can be merged before export.
With engines installed, recognition works offline. On a new machine, just download the models again.