New Multilingual Model Family Released by Bilibili

Bilibili Inc. has announced the release of Index-Translate, a new family of multilingual translation models built upon the Qwen3.5 architecture. According to documentation hosted on GitHub, the model family is designed to handle a wide range of linguistic tasks, covering 150 different languages.

The Index-Translate suite is structured into several specialized components designed to address specific translation requirements, including terminology adherence, formatting constraints, and content preservation.

Specialized Model Capabilities

The release includes several distinct models within the Index-Translate family, each targeting a specific modality or use case:

  • Index-Translate: The core text model, capable of translating structured content and community-specific expressions.
  • Index-Echo: A model designed to produce translated subtitles or speech that is conditioned on the original speaker’s voice.
  • Index-Homura: A specialized model that adjusts translations to meet a specific target syllable count.
  • Index-NativeLong: A model intended for full-document translation, maintaining context across multiple passages.

Early User Feedback and Performance Observations

Following the release, preliminary feedback has emerged from users on the social media platform Reddit. Some users have reported positive results in specific language pairs, such as Japanese to English. In unverified user testing, one individual noted that the model’s English-to-Arabic translations appeared highly accurate, suggesting a more literal approach to translation compared to other open models.

However, some users have raised questions regarding the model’s scale and comparative performance. One commenter noted that they had previously utilized Hy-MT1.5 and Hy-MT2 (1.8B) models, suggesting that the utility of Index-Translate would depend on the specific parameter size of the version being deployed.

While some users have expressed interest in testing the model in conjunction with other technologies, such as Whisper v3 Large for live subtitling, independent benchmarks have not yet been published to confirm the model’s performance against industry standards.

Developer Engagement

Bilibili representatives have actively engaged with the community on Reddit to solicit technical feedback. Developers have requested specific examples where the model may have fallen short, specifically regarding accuracy, natural dialogue, or tone, in an effort to identify areas for future improvement.