Bilibili Inc. has announced the release of Index-Translate, a suite of large language models (LLMs) designed to handle text, document, and video translation. According to a technical announcement shared by the Bilibili Index LLM team on Reddit, the new toolkit aims to address complex localization needs through specialized models for various media formats.

Technical Specifications and Model Variants

The core Index-Translate models support 150 text languages and are available in three parameter sizes: 2B, 9B, and a 35B-A3B preview version. According to the developers, the models allow users to specify particular terminology, writing styles, and output formats. This includes the ability to maintain consistency for product names or preserve technical structures such as JSON and placeholders during localization processes.

The release includes several specialized companion models tailored for specific workflows:

  • Index-NativeLong: Designed for whole-document translation, utilizing context to maintain consistency in names and terminology across long passages.
  • Index-Homura: A model focused on dubbing scripts that allows users to set a “syllable budget” for translated lines to ensure they fit the timing of the original audio.
  • Index-Echo: A model intended for generating multilingual subtitles or performing speech-to-speech translation by using the original speaker’s voice as a reference.

While the text models support 150 languages, the developers noted that Index-Echo currently supports a smaller subset of language pairs.

Availability and Licensing

The code and released model weights for the Index-Translate suite have been made available under the Apache-2.0 license. Users can access the models via GitHub and Hugging Face, with a functional demo hosted on the Bilibili Index website.

In a demonstration video shared by the team, the capabilities of the models were showcased through English-to-Japanese dubbing and an English video clip featuring subtitles in six different languages.

Industry Context

The release of these models targets several sectors, including video content creation, game localization, and professional document translation. While the technical details have been shared directly by the development team, independent third-party benchmarks were not included in the initial announcement to verify the performance claims across all 150 languages.