Bilibili Open-Sources Index-Translate Model Family Supporting 150 Languages
Bilibili's Index LLM team has open-sourced Index-Translate, a family of multilingual translation models built on Qwen3.5 with parameters ranging from 2B to 35B. The models support text translation across 150 languages along with specialized capabilities for audio dubbing, syllable length control, and long-document processing. By addressing specialized media localization challenges such as video audio syncing, internet slang adaptation, and long-context terminology consistency, Index-Translate pushes open-source AI translation tools forward. It provides content creators and developers with tailored models for complex multimedia translation workflows. The model suite includes Index-Translate for slang and structured content, Index-Echo for voice-cloned subtitles and dubbing, Index-Homura for syllable-controlled translation length, and Index-NativeLong for maintaining consistent character names across contexts up to 32K tokens. Weights for 2B, 9B, and 35B-A3B (preview) versions are available on Hugging Face and ModelScope.
## BACKGROUND
Machine translation based on general-purpose LLMs often struggles with localized slang, video dubbing syllable constraints, and term drift over long documents. Bilibili, a major Chinese video platform, previously released open-source lightweight models such as Index-1.9B to tackle long-context understanding.