Arc — a privacy-first end-to-end-encrypted messenger — uses TranslateGemma 4B as the on-device translation engine ("Rosetta") via the flutter_gemma package + MediaPipe LLM Inference. The upstream litert-community repository is gated (requires a Hugging Face login + Gemma license acknowledgement per user). Distributing the model through this mirror lets Arc download the model from end users' devices without bundling a Hugging Face access token in the Arc app or in app-config Firestore.

By downloading the model file from this mirror, you implicitly accept the upstream Gemma license terms. Atlas Associates Inc. acts only as a distribution mirror; the model weights are the property of Google.

Attribution

Item	Source
Original model weights	Google (`google/translategemma-4b-it`)
LiteRT-LM `.task` packaging	Google AI Edge / LiteRT community (`litert-community/TranslateGemma-4B-IT`)
License	Gemma Terms of Use
Mirror operator	Atlas Associates Inc. (this repository)

Files

File	Size	Format	Notes
`translategemma-4b-it-int8-web.task`	~3.63 GiB	MediaPipe LiteRT `.task` (int8, GPU/CPU)	Verbatim copy of the upstream file (`sha256` matches upstream).

Usage

This file is consumed by the flutter_gemma package via:

dart

await FlutterGemma.installModel(modelType: ModelType.gemmaIt)
    .fromNetwork(
      'https://huggingface.co/atlas-associates/TranslateGemma-4B-IT/resolve/main/translategemma-4b-it-int8-web.task',
      foreground: true,
    )
    .install();

The model runs entirely on-device — no data is sent to any server during inference. See Rosetta on-device translation in Arc for the user-facing product context.

License

Gemma — see https://ai.google.dev/gemma/terms and the LICENSE file in this repository.

By downloading, you agree to the Gemma Terms of Use as specified by Google. The terms include responsible-use commitments and prohibited-use limitations that you must read and comply with.

Contact

Issues with the mirror itself (broken file, missing version, takedown requests):

Atlas Associates Inc. contact
Open an issue on GitHub: Atlas-Associates-Inc/Arc-V2

Issues with the model behavior (translation quality, performance):

Refer upstream: google/translategemma-4b-it and litert-community/TranslateGemma-4B-IT.