On April 29, the Tencent Huan Yuan team officially announced the open-source release of its ultra-quantized compressed translation model - Hy-MT1.5-1.8B-1.25bit. The biggest highlight of this model is that it compresses the translation capabilities for 33 languages to about 440MB, which means even mobile devices with limited storage space can achieve high-quality real-time translation without an internet connection.
Ultra Compression: A "Weight Loss" Revolution for Mobile Memory
This achievement stems from Tencent's previously released professional translation large model Hy-mt1.5. The original 1.8B model requires about 3.3GB of memory at normal precision, which is a heavy burden for mobile device operation. To solve this pain point, the research team adopted extreme quantization compression technology, reducing the parameter expression from 16 bits (16-bit) to as low as 1.25 bits.
In simple terms, this process is similar to compressing a 4K high-definition large image into a micro image with a tiny volume without losing key details. For different performance devices, Tencent has also launched two quantization schemes, 2-bit and 1.25-bit, ensuring that the model maintains excellent semantic understanding ability after "weight loss."
Performance Testing: Translation Quality Exceeds Mainstream Competitors Offline
Although the size has been significantly reduced, the performance has not declined. According to official evaluation data, this lightweight model with 1.8B parameters has achieved translation quality comparable to or even surpassing mainstream commercial systems like Google Translate in multiple benchmark tests, and even in some dimensions, it can compete with large models with 235B parameters.
Currently, the model supports 33 languages including Chinese, English, Japanese, French, Russian, and Arabic, as well as minority languages such as Tibetan and Mongolian. It also supports five dialects and bilingual translation, with a total of 1056 translation directions, greatly expanding the application boundaries of offline translation.
Deep Integration: Prioritizing Privacy Security and Convenient Experience
Beyond just weight open-source, this technology has shown great practical value in real scenarios. In the latest adapted demonstration version, the model supports a "background word selection mode," allowing the translation function to be available on demand regardless of whether users are reading emails offline or browsing local web pages.
Notably, since the translation process runs entirely on the local device without uploading any personal privacy information or collecting data from the cloud, it provides reliable protection for users with high requirements for data security. Currently, this translation capability has been applied in several core business scenarios within Tencent, including meeting systems, office software, and browsers.
To facilitate developers and tech enthusiasts to experience, the relevant models have been synchronized and launched on Huggingface and Moba Community. This marks that high-precision translation technology is accelerating from the cloud to the terminal, making translation services truly become a basic tool that is always at hand.
