Bilibili Releases Index-Translate, a Translation Model Supporting 150 Languages

The Bilibili Index team released the multilingual translation model family Index-Translate on September 30, 2026. Based on Qwen3.5, these models support 150 languages and consist of five models for various purposes, including text, speech, and long-form translation. All model weights and code are distributed under the Apache License 2.0, allowing for commercial use.

Model Composition and Features

• Index-Translate (2B/9B/35B-A3B Preview): Integrates support for text translation, instruction following, and meme translation.
• Index-Echo (2B/9B): A speech translation model supporting speech-to-text (S2TT) and speech-to-speech (S2ST) conversion. It uses a Hidden2CV mapper that directly maps hidden states instead of a cascading approach to preserve voice timbre.
• Index-Homura (2B/9B): A dubbing translation model that controls the target syllable count. On the SandGlass benchmark, the 9B model achieves a rate of 81.92% for generating outputs within 10% of the target syllable count.
• Index-NativeLong (2B/9B): A long-form document translation model that processes entire documents at once to maintain terminology consistency. It recorded a COMET score of 0.7382 at a 64K token length, preventing the performance drop seen in other models.

Performance and Limitations

The text model achieved a score of 0.8794 on FLORES-200, showing the best performance among comparison models. However, due to translation specialization, general capabilities (C-Eval, GPQA, etc.) declined compared to the base model. Speech translation showed higher translation quality than Gemini-3.1-Pro, but the long-form translation evaluation set was self-created, raising the possibility of overfitting. The 35B-A3B model is a preview version currently under training.