# Third-party components Babelglass loads all inference assets from local files. Model download code exists only in `scripts/prepare-model.py`, which is not part of the running lens. - RapidOCR 1.4.4: [project](https://github.com/RapidAI/RapidOCR), Apache-2.0. Its wheel includes PaddleOCR detection and orientation models. - English, German, Spanish, French, Portuguese and Italian recognition: [RapidOCR Latin PP-OCRv5 ONNX model](https://www.modelscope.cn/models/RapidAI/RapidOCR/files?Revision=v3.9.2), Apache-2.0. The embedded character dictionary includes UI punctuation, German umlauts and ß, Spanish accents, ñ, French accents and inverted punctuation. Provenance and SHA256 are retained in `models/ocr/en-source.json`, `de-source.json`, `es-source.json`, `fr-source.json`, `pt-source.json` and `it-source.json`. - Russian recognition: [RapidOCR Cyrillic PP-OCRv5 ONNX model](https://www.modelscope.cn/models/RapidAI/RapidOCR/files?Revision=v3.9.2), Apache-2.0. The character dictionary is embedded in the model and includes Russian Ё/ё. Provenance and pinned SHA256 are retained in `models/ocr/ru-source.json`. - Simplified and Traditional Chinese recognition: [RapidOCR Chinese PP-OCRv5 ONNX model](https://www.modelscope.cn/models/RapidAI/RapidOCR/files?Revision=v3.10.0), Apache-2.0. The embedded dictionary covers both scripts. Provenance and pinned SHA256 are retained in `models/ocr/zh-source.json` and `zt-source.json`. - Translation: [Argos English–German and German–English 1.3](https://data.argosopentech.com/argospm/v1/), preconverted CTranslate2 weights and SentencePiece tokenizer. The publisher's metadata, README and archive SHA256 are retained in `models/en-de/` and `models/de-en/`. Argos maintainer PJ-Finlay confirmed that `.argosmodel` binaries use the same MIT/CC0 license as Argos Translate in [the publisher's licensing clarification](https://github.com/argosopentech/argos-translate/issues/533#issuecomment-5160080718). The publisher's MIT notice is retained in `licenses/argos-MIT.txt`. - Spanish translation: [Argos English–Spanish and Spanish–English 1.0](https://data.argosopentech.com/argospm/v1/), preconverted CTranslate2 weights and SentencePiece tokenizers. The publisher's metadata, README and pinned archive SHA256 are retained in `models/en-es/` and `models/es-en/`. These SentencePiece packages work with the existing local inference engine. The same publisher MIT/CC0 clarification above applies; retained README files preserve the training-data and Stanza attribution. - Russian translation: [Argos English–Russian and Russian–English 1.9](https://data.argosopentech.com/argospm/v1/), preconverted CTranslate2 weights and SentencePiece tokenizers. The publisher's metadata, README and pinned archive SHA256 are retained in `models/en-ru/` and `models/ru-en/`. The same publisher MIT/CC0 clarification above applies; retained README files preserve training-data and Stanza attribution. - French translation: [Argos English–French and French–English 1.9](https://data.argosopentech.com/argospm/v1/), preconverted OPUS-MT CTranslate2 weights and SentencePiece tokenizers. Package metadata, README attribution and archive SHA256 are retained in `models/en-fr/` and `models/fr-en/`. The package READMEs attribute the underlying OPUS-MT models to Jörg Tiedemann and Santhosh Thottingal, “OPUS-MT — Building open translation services for the World”, EAMT 2020, Lisbon, and specify [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) for the original models. Their preconverted packaging is supplied by Argos. - Portuguese translation: [Argos English–Portuguese and Portuguese–English 1.9](https://data.argosopentech.com/argospm/v1/), preconverted OPUS-MT weights and SentencePiece tokenizers. The general `pt` packages are bundled; `pb` is a separate Brazilian Portuguese package. Metadata, README attribution and archive SHA256 are retained in `models/en-pt/` and `models/pt-en/`. The package READMEs attribute the underlying OPUS-MT models to Jörg Tiedemann and Santhosh Thottingal, “OPUS-MT — Building open translation services for the World”, EAMT 2020, Lisbon, and specify [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) for the original models. Their preconverted packaging is supplied by Argos. - Italian translation: [Argos English–Italian and Italian–English 1.0](https://data.argosopentech.com/argospm/v1/), preconverted CTranslate2 weights and SentencePiece tokenizers. Metadata, README and pinned archive SHA256 are retained in `models/en-it/` and `models/it-en/`. The same publisher MIT/CC0 clarification above applies; package READMEs retain the OPUS training-data and Stanza attribution. - CTranslate2: [project](https://github.com/OpenNMT/CTranslate2), MIT. - SentencePiece: [project](https://github.com/google/sentencepiece), Apache-2.0. - ONNX Runtime: [project](https://github.com/microsoft/onnxruntime), MIT. - Simplified and Traditional Chinese translation: [Argos English–Chinese packs 1.9](https://data.argosopentech.com/argospm/v1/), separate preconverted OPUS-MT CTranslate2 weights and SentencePiece tokenizers for all four directions. Metadata, publisher README attribution and archive SHA256 are retained in `models/en-zh/`, `zh-en/`, `en-zt/` and `zt-en/`. Their READMEs credit Jörg Tiedemann and Santhosh Thottingal, “OPUS-MT — Building open translation services for the World”, EAMT 2020, Lisbon, and specify CC BY 4.0 for the underlying models. These notices are retained in offline bundles. - Pyphen 0.17.2: [project](https://github.com/Kozea/Pyphen), GPL-2.0-or-later, LGPL-2.1-or-later or MPL-1.1. Its bundled LibreOffice hyphenation dictionaries retain their individual license/readme files in `pyphen/dictionaries/`. - Python: Python Software Foundation license. Bundles retain the standard-library license and dependency distribution metadata. Other bundled dependencies and their version pins are in `requirements.lock`; their installed `.dist-info` directories retain supplied license files. GNOME Shell, Mutter, GJS and system libraries are supplied by the desktop OS. The Windows frontend uses .NET/WPF and the Windows SDK C#/WinRT projection; the self-contained publish retains supplied distribution notices. .NET is licensed under MIT: https://github.com/dotnet/runtime and https://github.com/dotnet/wpf. C#/WinRT: https://github.com/microsoft/CsWinRT. Windows bundles also include colorama 0.4.6 (BSD-3-Clause), with its installed distribution metadata. The Windows API bindings are authored for this project. App-local Microsoft Visual C++ runtime DLLs must come from a distribution with appropriate redistribution rights; preserve the corresponding runtime terms. Standalone Python runtime provenance is recorded in the Windows bundle manifest.