Bilibili open-sources Index-Translate translation model for 150 languages

On October 6, 2026, a member of Bilibili's Index LLM team posted on V2EX about Index-Translate, a translation model the team recently open-sourced. The model is trained on Qwen3.5 Base and covers 150 text languages. It ships in three versions: 2B, 9B and 35B-A3B preview, and a web demo is available at index-translate.bilibili.com for people who want to try it directly.

The same post announced a free API that is compatible with the OpenAI interface, so users can hook it up simply by changing the base URL of an existing client. The team also included a Python calling script with no external dependencies, hosted at inference/llm/call_api.py in the bilibili/Index-Translate repository on GitHub.

According to the post, the team trained specifically for instruction following in addition to translation quality. Users can pass in a glossary, specify a writing style, or ask the model to preserve content and formatting — for example, keeping character names and cultivation-realm terms consistent when translating a novel, or instructing the model to output only the translated text without explanations.

The post rounds off with several branches built alongside the main model: Echo, which turns speech directly into subtitles or performs end-to-end speech translation with voice cloning; Homura, which lets users set a syllable budget for the output, aimed at cases where timing is constrained; and Nailong, a long-document translation model intended to reduce chunking. The available material stops there — no further details on training data, evaluation results or future plans were given.