Index-Translate

Index-Translate is a family of AI translation models from Bilibili’s Index team. The text models translate across 150 languages and accept instructions about terminology, style and formatting. The team released 2B, 9B and 35B-A3B preview weights on 30 September 2026, followed by compressed versions on 3 October.

A glossary can guide the translation

The models are built on Qwen3.5. A user can supply a glossary, ask for a particular tone or require the output to preserve code tags and document structure. Those controls are useful when a product name or technical term has to stay consistent across a document.

The 35B-A3B preview model card describes a model with about 35 billion parameters and roughly 3 billion active for each token. It uses the Apache 2.0 licence. Its configuration supports 262,144 tokens, while the example serving setup uses a smaller 32,768-token limit; the actual service configuration determines the usable length.

Different family members handle different jobs

The project also includes Index-Echo for translated speech or subtitles, Index-Homura for translations with a target syllable count, and Index-NativeLong for whole documents. The public demonstration and free API provide another way to try the text model.

The published evaluation is the team’s own. Results vary by language and task, and the model card identifies low-resource languages and combined instructions as harder cases. For an important translation, check names, numbers and specialised terms against the source text.