Bilibili's Index LLM team has officially released the Index-Translate model family, which is built upon the Qwen3.5 architecture. The project includes text models with 2B, 9B, and 35B-A3B (preview) parameters, which are now available on Hugging Face and ModelScope.
The new model family supports 150 languages, including Chinese and English. It is designed to handle specific translation instructions such as terminology management, formatting, and content preservation. The system also extends its capabilities to voice translation, syllable-controlled output, and long-document processing.
The Index-Echo component enables the generation of subtitles or dubbing in target languages. When performing dubbing, the model utilizes voice characteristics from the source audio to maintain consistency. The team demonstrated the model's ability to adjust phrasing based on target length requirements and handle long-form text contexts of approximately 32K tokens.
