Bilargo Translate — AI & Model Notice
Effective date: 2026-07-07
This notice describes the AI models used by Bilargo Translate (the "App") and related privacy and limitation information. It supplements our Privacy Policy and Terms of Use.
1. Local AI translation
The App uses local AI models for machine translation. By default, all inference runs on your Mac using Apple's MLX framework — your source text, translated text, and translation history are processed locally and are not sent to a cloud AI service unless you explicitly configure an optional cloud backend with your own API key (see Section 2).
2. Model catalog
TranslateGemma
The translation models are from the TranslateGemma family, provided by Google and converted to MLX format by the mlx-community:
- TranslateGemma 4B (4-bit / 8-bit)
- TranslateGemma 12B (4-bit / 8-bit)
These models are purpose-built for machine translation; the app exposes 49 of their languages in the picker.
Gemma utility model
A compact instruction-tuned Gemma model (E4B) serves as a local utility model for auxiliary features: glossary refinement, text improvement ("Improve" mode), and synonym suggestions. It is not the primary translation engine — TranslateGemma is. The same Gemma Terms of Use apply.
On-device text recognition (OCR)
When you translate an image or a scanned PDF, text recognition runs entirely on your Mac. The primary OCR engine is Apple's built-in Vision framework. For languages Vision does not recognize (for example Polish and several other scripts), the App falls back to Tesseract OCR (Apache License 2.0). The Tesseract language data (tessdata) it uses is bundled inside the App — no OCR model is downloaded, and nothing about OCR contacts the network. Images and the text extracted from them are processed locally and never leave your device.
Optional backends you configure yourself
Beyond the built-in local models, the App can connect to backends you provide, at two distinct privacy levels:
- Local server on your own network (for example Ollama or LM Studio, on this Mac or another machine you control): your text travels only within your own infrastructure. Any model you host there is governed by that model's own license.
- Cloud API with your own key (for example OpenAI, Anthropic, Mistral, Google, or OpenRouter): your text is sent to the selected provider under that provider's own terms and privacy policy. Cloud backends are labeled "CLOUD" in the App and are disabled in offline mode. See Privacy Policy, Section 4a.
3. Model sources and licenses
| Model family | License / Terms | Links |
|---|---|---|
| TranslateGemma | Gemma Terms of Use + Prohibited Use Policy | Gemma Terms · Prohibited Use Policy |
| MLX community conversions | Original model license applies; conversion tooling under MIT | mlx-community on Hugging Face |
| Tesseract OCR + Leptonica (bundled OCR fallback) | Apache License 2.0 | Tesseract · Leptonica |
Users who download or use TranslateGemma models must comply with the Gemma Terms of Use and the Gemma Prohibited Use Policy.
4. Privacy: how models interact with your data
- Source text, translated text, and translation history are processed locally on your Mac. We do not intentionally send this content to our servers or to any cloud AI inference service.
- Optional model downloads contact a third-party model host (such as Hugging Face) and may expose standard network metadata, including your IP address, the requested model repository, and timestamps. After the download completes, translation runs fully offline.
- Support communications: if you contact support and voluntarily attach logs or text, that content may leave your device. We recommend not sharing sensitive text unless necessary.
- Apple / App Store flows: standard App Store interactions (purchase, restore, app updates) involve Apple infrastructure under Apple's own privacy policy.
5. Limitations and disclaimers
- Translation errors and hallucinations: AI models may produce inaccurate, incomplete, misleading, or contextually inappropriate translations. Output is provided "as is."
- Not a certified or sworn translation: the App does not produce legally recognized translations.
- No legal, medical, or financial advice: translation output must not be relied upon for critical decisions in legal, medical, financial, safety-related, or regulatory contexts.
- User verification required: you must independently verify any translation output before using it for critical purposes.
6. Model availability
Model availability may change for legal, licensing, security, or technical reasons. We may add, update, or remove support for any model at our discretion, subject to your rights under applicable consumer law and the Terms of Use.
7. Contact
For questions about AI models or this notice: