Bilibili, the Chinese video and streaming platform, has released a family of open-weight translation models called Index-Translate. Its repository dates the release to 30 September, with weights for 2B, 9B and a preview 35B-A3B version published under an Apache-2.0 licence. The models cover 150 languages, and the 35B configuration is a mixture-of-experts design with 35 billion total parameters and roughly 3 billion active.
Alongside the text models the team lists speech translation, syllable-controlled translation and long-document variants. That reads less like a research release and more like the internal toolkit of a company that subtitles and dubs a very large amount of video, published rather than kept.
On the numbers, every figure available so far is Bilibili’s own and none has been independently reproduced. The repository reports a FLORES score of 0.8794 for the 35B-A3B model and 0.8789 for the 9B, close enough that the smaller model looks like the sensible default. On WMT26 it lists 76.76 and 75.35 respectively. The 35B weights are labelled a preview, with an official version still on the to-do list, so treat the top-line scores as provisional.
Why it matters: translation is one of the areas where open weights genuinely compete with paid APIs, and a permissive licence on a 150-language model is useful to anyone who cannot send text to a third party.
Source: the Index-Translate repository.