About CranScan

CranScan — a private, offline document & receipt scanner (Cyrillic + Latin OCR, on-device AI, spending analytics). Everything runs on your device; nothing leaves your phone.

CranScan is a closed-source app published by @samoylenkodmitry. This repository is not the app — it is only the open host for CranScan's optional downloadable model packs (the OCR language packs and Receipt-AI vision models the app can fetch on demand). See Model packs below.

• Get it on Google Play: https://play.google.com/store/apps/details?id=com.cranscan.app
• Download on the App Store: https://apps.apple.com/app/id6795579921
• Website: https://cranscan.dmitrysamoylenko.in
• Privacy policy: https://cranscan.dmitrysamoylenko.in/privacy/

Screenshots

_{App UI, v1.0.0 (Android, iPhone and iPad). Point your camera at a page or receipt — CranScan crops it, reads the text on-device, extracts merchant/date/total/line items, and keeps a searchable, private library with spending analytics.}

---

Model packs

Downloadable model packs for CranScan: OCR language packs (packs-v1) and Receipt-AI vision models (ai-v1). These are optional — the app ships with a bundled Cyrillic + Latin recognizer and only downloads a pack when you choose to.

Each OCR pack is a ZIP containing rec.rten (a PaddleOCR recognition model converted to the rten runtime) and dict.txt. Install in-app via Settings → Text recognition (one-tap download or manual import).

OCR language packs (release packs-v1)

Pack	Model	Size	Scripts / languages
cyrillic.zip	cyrillic_PP-OCRv5_mobile_rec	7 MB	Cyrillic incl. Serbian (ЂЈЉЊЋЏ), Bulgarian, Russian, Ukrainian, Belarusian + Latin digits
latin.zip	latin_PP-OCRv5_mobile_rec	7 MB	Latin (Western/Central European)
greek.zip	el_PP-OCRv5_mobile_rec	7 MB	Greek
arabic.zip	arabic_PP-OCRv5_mobile_rec	8 MB	Arabic
hindi.zip	devanagari_PP-OCRv5_mobile_rec	8 MB	Devanagari
korean.zip	korean_PP-OCRv5_mobile_rec	8 MB	Korean
thai.zip	th_PP-OCRv5_mobile_rec	7 MB	Thai
tamil.zip	ta_PP-OCRv5_mobile_rec	8 MB	Tamil
telugu.zip	te_PP-OCRv5_mobile_rec	8 MB	Telugu
chinese.zip	PP-OCRv5_server_rec	76 MB	Chinese (simpl.+trad.), English, Japanese, Pinyin — server tier
multi-small.zip	PP-OCRv6_small_rec	19 MB	Chinese, English, Japanese, Greek, Latin-script languages — beats PP-OCRv5 server on most benchmarks at a quarter of the size
multi-hq.zip	PP-OCRv6_medium_rec	67 MB	Chinese, English, Japanese, Greek, Latin-script languages — highest accuracy available (+5.1% over PP-OCRv5 server)

Note: no server/large-tier Cyrillic recognition model exists in the PaddleOCR family yet (PP-OCRv6 covers 50 languages but not Cyrillic); cyrillic.zip (mobile tier) is the best available for Serbian/Bulgarian and is the only pack whose dictionary includes the Serbian-specific letters ЂЈЉЊЋЏ.

Receipt AI packs (release ai-v1)

Local vision-language models that read a receipt photo directly — merchant, date, total, currency, line items and a spending category — far more robustly than OCR + heuristics. Used by the CranScan desktop app (Settings → Receipt AI); everything runs on-device via a bundled llama.cpp llama-server (build b9873, CPU).

Pack	Model	Download	RAM while reading	Notes
qwen3vl-4b	Qwen3-VL-4B-Instruct Q4_K_M	~3.3 GB (2 shards + mmproj)	~6.7 GB	Best quality; recommended
qwen3vl-2b	Qwen3-VL-2B-Instruct Q8_0	~2.3 GB	~4.5 GB	Faster, good on clean receipts

Each pack is described by a <id>.manifest.json release asset:

{
  "id": "qwen3vl-4b",
  "ram_hint_mb": 6700,
  "total_bytes": 3333463168,
  "model": "Qwen3-VL-4B-Instruct-Q4_K_M-00001-of-00002.gguf",
  "mmproj": "mmproj-Qwen3-VL-4B-Instruct-F16.gguf",
  "files":   [ { "name": "…", "bytes": 123, "sha256": "…" } ],
  "servers": { "linux-x86_64": { "name": "…tar.gz", "bytes": 123, "sha256": "…", "bin": "llama-server" } }
}

• files lists the GGUF shards + multimodal projector (every file sha256-pinned; models over 2 GB are split with llama-gguf-split, llama-server loads the first shard and finds the rest).
• servers maps a platform (linux-x86_64, macos-arm64, macos-x86_64, windows-x86_64) to a minimal, flat llama-server archive (binary + required shared libraries, repacked from the official llama.cpp b9873 release).
• The app downloads files + its platform's server archive, verifies every sha256, unpacks the server and talks to it over localhost — nothing leaves the device.

Models are the official Qwen3-VL instruct releases (Apache-2.0), GGUF quantizations from the ggml-org/unsloth conversions; llama.cpp is MIT.

Detection model

paddle-det-v6-medium.rten — PP-OCRv6_medium_det (62 MB), a server-grade text detector (+4.6% Hmean over PP-OCRv5 server det). Not an installable pack; intended to be bundled with the app in place of the default PP-OCRv6 small detector for far better line detection on crumpled/hand-held receipts.

Sources

Models originate from PaddleOCR (Apache-2.0). ONNX exports: official PaddlePaddle Hugging Face *_onnx repos (cyrillic_PP-OCRv5_mobile_rec_onnx, PP-OCRv6_medium_rec_onnx, PP-OCRv6_small_rec_onnx, PP-OCRv6_medium_det_onnx) and monkt/paddleocr-onnx (the PP-OCRv5 mobile language packs and PP-OCRv5 server chinese). Converted with rten-convert; dictionaries extracted from each model's inference.yml (PostProcess.character_dict). The cyrillic ONNX ships as opset 7 and must be upgraded (e.g. onnx.version_converter.convert_version(model, 13)) before rten-convert, otherwise Slice ops fail at runtime.

---

CranScan is closed-source software © Dmitry Samoylenko. The models hosted here retain their upstream licenses (PaddleOCR / Qwen3-VL: Apache-2.0; llama.cpp: MIT).