1. Abstract
Hy-MT1.5-1.8B-1.25bit is a low-bit offline translation model based on HY-MT1.5-1.8B launched by Tencent's hybrid team, with a parameter size of 1.8B and a quantized volume of about 440MB. It is deployed for mobile and edge devices, supporting 33 languages, 5 dialects/minority languages, and 1056 translation directions. According to the official technical report, the model has strong parameter efficiency in standard Chinese, foreign and English-foreign translation tasks, and can compete with some commercial translation APIs and larger-scale open source models. It is still recommended to review the specific effect according to the business language and test set.
2. Core features
- Extremely low volume: Compress the FP16 approximately 3.3GB model to approximately 440MB through Sherry 1.25-bit ternary quantization.
- Mobile offline: The official Android Demo APK is provided to experience offline translation in a network-free environment.
- Multilingual coverage: Support Chinese, English, Japanese, Korean, French, Spanish, Tibetan, Mongolian, Uyghur, Cantonese, etc.
- Translation Capabilities Enhancement: Support tasks such as terminology intervention, contextual translation, and format maintenance.
- Compression framework support: Relying on the AngelSlim toolchain, it covers quantization, deployment, and subsequent compression algorithm iterations.
3. Installation
- Android experience: Download the official Demo APK and experience offline translation directly after installation.
- Model weights: Download the 1.25-bit weights or GGUF version from Hugging Face.
- AngelSlim toolchain: can be installed by pip install angelslim; The source code method can be cloned from GitHub and then run python setup.py install.
- Note: The official documentation shows that the STQ kernel related to llama.cpp is still in progress, and the project deployment must be subject to the latest repository description.
4. Typical use cases
- Mobile offline translation: suitable for travel, cross-border communication, weak or no network scenarios.
- Privacy-sensitive text translation: Text can be processed locally, reducing the need to upload external services.
- Mutual translation of minor languages: covering Tibetan, Mongolian, Uyghur and other ethnic minority languages.
- Enterprise terminology translation: It can be combined with terminology intervention templates to improve the consistency of professional vocabulary.
- Cross-app word picking: The Android demo demonstrates the background word picking mode, which is suitable for translating emails, web pages, and chat content.
5. Ecology and competing products
- Ecosystem: The project also provides Hugging Face weights, GGUF version, Android demo, AngelSlim documentation, and GitHub repositories.
- Competing products: It can be compared with commercial services such as Google Translate, Microsoft Translator, DeepL, and Doubao Translation, as well as Tower-Plus, Qwen, and other open source MT models.
- Differences: The core advantages of Hy-MT1.5-1.8B-1.25bit are its small size, offline accessibility, and wide language coverage, making it suitable for end-side translation rather than only for cloud APIs.
6. Limitations and precautions
- The official benchmark is not the same as all real business scenarios, and needs to be retested according to the direction, field and text length.
- Low-bit quantization may produce unstable results in complex long sentences, technical terms, or low-resource language.
- Android APK is suitable for experience, and production deployment still needs to pay attention to permissions, update mechanisms, and security audits.
- GGUF and the end-side kernel ecosystem are still iterating, and the current inference backend support should be confirmed before deployment.
- When involving high-risk texts such as law, medical care, and contracts, manual review is recommended.
7. Project address
https://github.com/tencent/AngelSlim
8. Frequently asked questions
Q: Can Hy-MT1.5-1.8B-1.25bit be translated completely offline?
A: The official Android demo is for offline translation experience, and the model can be run locally after downloading; For actual deployment, please refer to the latest documentation.
Q: What languages does Hy-MT1.5-1.8B-1.25bit support?
A: There are 33 official languages and 5 types of dialects/minority languages, including Chinese, English, Japanese, Korean, Tibetan, Mongolian, Uyghur, Cantonese, etc.
Q: Which is better, Hy-MT1.5-1.8B-1.25bit or Google Translate?
A: The official report highlights its strong competitiveness on several standard benchmarks; However, Google Translate is a continuously updated commercial service, and it is recommended to test it according to specific language directions and domain text.
Q: Does 1.25-bit quantization lose translation quality?
A: Officially, Sherry quantifies near-lossless performance while compressing volume; The actual effect still depends on the language direction and the input content.
