Bilibili Releases Open-Source Index-Translate Models

Bilibili's Index LLM team (IndexTeam) has released the open-source Index-Translate machine translation model family under the Apache-2.0 license. Based on Qwen3.5, the models translate between 150 languages, including Russian, and follow translation instructions.

Compiled by Sergey KostenchukPublished 2026-10-01Updated 2026-10-01

2026-10-01 Research HuggingFace

🤖 Bilibili Releases Open-Source Index-Translate Models

Bilibili's Index LLM team (IndexTeam) has released the open-source Index-Translate model family based on Qwen3.5: 2B, 9B, and MoE 35B-A3B under the Apache-2.0 license (flagship in preview). The models work with 150 languages, including Russian, maintain glossaries, markup, and placeholders, and understand Chinese meme slang.

🌍 The open 9B model outperforms API flagships GPT-5.6-Sol and DeepSeek-V4.1-Flash on COMET-22 for FLORES-200—primarily due to instruction following, which is critical for subtitles and localization. In WMT26, judged by GPT-5.6-Sol, it still lags behind.

👤 The models are already available on Hugging Face: the 2B model runs on approximately 8 GB of VRAM, the 9B model on 24 GB, and Russian is supported in both directions. Glossary and markup preservation will be useful for subtitles and documentation.

Source 1: https://index-translate.bilibili.com/ Source 2: https://huggingface.co/IndexTeam/Index-Translate-9B

Sources