AI News (2026/4/30): Tencent Hunyuan Open-Sources Offline Translation Model for Mobile Devices: Hy-MT1.5-1.8B-1.25bit
Executive Summary:
Tencent Hunyuan has open-sourced the offline translation model Hy-MT1.5-1.8B-1.25bit for mobile devices, supporting 33 languages and 1056 translation directions. The model is compressed to 440MB using Sherry sparse quantization technology and can run locally on a mobile device without an internet connection. The translation quality surpasses mainstream systems like Google Translate.
Tencent Hunyuan Open-Sources Offline Translation Model for Mobile Devices: Hy-MT1.5-1.8B-1.25bit
Details
Tencent Hunyuan recently announced the open-sourcing of its mobile offline translation model, Hy-MT1.5-1.8B-1.25bit. This model supports 33 languages and 1056 translation directions, and through advanced sparse quantization technology, it has been compressed to 440MB, ensuring it can run locally on a mobile device without an internet connection. Hy-MT1.5-1.8B-1.25bit is built on the foundation of the 1.8B parameter Hy-MT1.5, significantly improving translation quality.
Key Points
[33 Languages]: Supports 33 languages, including English, Chinese, Japanese, Korean, and other major global languages.
[1056 Translation Directions]: Provides bidirectional translation between any supported language and the other 32 languages, totaling 1056 translation directions.
[440MB]: Using Sherry sparse quantization technology, the 1.8B parameter model is compressed to 440MB, significantly reducing storage and computational resource requirements.
[Sherry Sparse Quantization Technology]: This technology, accepted by ACL 2026, effectively reduces the model size and computational complexity by combining sparsification and quantization.
[No Internet Connection Required]: Users can perform real-time translations using the model without an internet connection, enhancing flexibility and reliability in various usage scenarios.
[Translation Quality]: In multiple benchmark tests, the translation quality of Hy-MT1.5-1.8B-1.25bit surpasses mainstream systems like Google Translate.
AI-ALL In-Depth Commentary
The open-sourcing of the Hy-MT1.5-1.8B-1.25bit model by Tencent Hunyuan represents a significant technological breakthrough and demonstrates strong potential in practical applications. By compressing the large model to 440MB and supporting 33 languages and 1056 translation directions, this model provides a new solution for offline translation on mobile devices. The application of Sherry sparse quantization technology not only reduces the demand for storage and computational resources but also maintains high translation quality, offering important insights for deploying AI models in resource-constrained environments. Additionally, the feature of not requiring an internet connection makes the model particularly advantageous in remote areas or under unstable network conditions, potentially further driving the adoption and application of AI translation technology.