Bilibili’s Index-Translate: Model Sizes, Languages and Specialized Branches

Bilibili’s Index-Translate is a Qwen3.5-based family of translation models and tools for text, speech, controlled dubbing and long documents. Its text models come in 2B, 9B and 35B-A3B preview sizes and are described as supporting 150 languages, with instructions for handling terminology, formatting and content preservation. The other branches focus on specific translation challenges rather than extending every feature to every task. 25

Text models: three sizes, one translation focus

The family’s text releases are the 2B and 9B models, plus the 35B-A3B model marked as a preview. The latter is a mixture-of-experts model with 35 billion total parameters and about 3 billion active per token. The available descriptions say the text models support translation across 150 languages and can follow instructions about terminology, formatting, style, structure and other output requirements. They don’t establish that every model size performs identically across all language pairs. 245

For licensing, the Index-Translate 2B model card lists Apache-2.0. Release coverage describes the family as Apache-2.0 as well, but it’s sensible to check the license attached to the specific weights and files you plan to use. 49

What the specialized branches do

• Index-Echo handles speech translation. Its speech-to-text workflow creates timestamped subtitles for Chinese audio or video translated into English, Japanese or Spanish, with voice-activity segmentation, sliding context and an optional glossary. Its speech-to-speech branch is designed to use the source speaker’s voice characteristics in dubbing. Those are documented capabilities and goals—not a guarantee of identical voice quality in every output. 2123
• Index-Homura adjusts translated wording to a specified syllable count, a constraint intended for dubbing scripts where spoken length matters. Meeting a syllable target alone does not prove that a line will sound natural or fit a particular scene perfectly. 2527
• Index-NativeLong, also published under the name Index-Nailong, is built for document-level translation. It takes the full source document into account to help maintain continuity, terminology and references across longer text. The released document client provides fixed templates for Chinese–English and Chinese–Japanese translation, so its documented workflow is narrower than the text family’s 150-language coverage. 326

Tools, evidence and limits

The repository includes a browser extension for translating web pages with a locally deployed model exposed through an OpenAI-compatible API. For video dubbing, the documented workflow includes audio extraction, vocal separation, timestamped speech segmentation, translation and alignment back to the video timeline. These tools show how the models can be incorporated into translation workflows, while also making clear that dubbing involves processing and alignment steps beyond translation itself. 2022

The team’s examples and model materials cover different tasks, including text translation, subtitles, syllable-constrained output and long documents. However, examples and reported evaluations should not be treated as independent proof that one model is best across every language, task or use case. The 35B-A3B text release is explicitly a preview, and the team lists shareable NativeLongBench material as planned. The available evidence therefore supports a practical summary of what each branch is designed to do, not a blanket claim of consistent results or guaranteed voice, timing or document-wide accuracy. 526