Bilibili released Index-Translate, an open-weights translation model family

A video company has released an open-weights translation model family that handles text, images, speech and long documents, with the strongest version comparing itself to the DeepSeek and Gemini numbers it publishes. Bilibili's Index LLM Team released Index-Translate on 30 September under Apache-2.0, published quantized builds on 3 October, opened a free public API and benchmark page on 4 October, and added native multimodal image translation on 10 October. The repository stands at 1,505 stars.

The family is built on Qwen3.5 and covers 150 languages. Three sizes ship: a 2B model that runs on an ordinary workstation, a 9B model for a single GPU, and a 35B-A3B preview that the team calls the flagship. Each is published as weights, with GGUF, FP8 and FP4 quantized builds for teams running on consumer hardware.

The numbers the team publishes

Every score here comes from Bilibili's own benchmark page, not an independent test. On FLORES, the team reports 0.8794 for Index-Translate-35B-A3B against the 0.8762 it lists for DeepSeek V4.1 Flash and 0.8750 for Gemini 3.5 Flash Lite. The team's WMT26 score for 35B-A3B is 76.76. On instruction-following translation, the team reports 0.6901 quality and an IFscore of 0.8336. The MEME benchmark score is 0.7405.

For low-resource language pairs, which are where most paid APIs lose fluency, the team reports COMET-22 of 0.8168 on FLORES minor pairs and 0.5151 quality on instTrans_minor. The claim is that the Qwen3.5 base plus the team's own training data pulls the gap closed where incumbent APIs still fall off.

The multimodal side

Index-Translate ships with a native vision encoder for image translation, which the team describes as working without an OCR step: text in the image, layout and context are read by one model. A sibling model family called Index-Echo handles speech-to-subtitles and voice-cloned dubbing, Index-Homura does syllable-controlled dubbing, and Index-NativeLong is a long-document consistency model. Each sibling is published in 2B and 9B sizes.

Where it would matter

For a team that translates medical, legal or payment copy, the choice today is between sending every string to one of three paid APIs or running an older open model that is further from the frontier. The Apache-2.0 release is a third option: a model recent enough to compare to those APIs on the team's own benchmark, with weights that stay inside the team's own network and sizes that fit the hardware a working engineering team already has.

The limit is that every number cited here comes from Bilibili. Independent testing of the FLORES delta against DeepSeek and Gemini has not yet appeared. A team ready to adopt the model should run it on its own data and compare the output against whatever paid API it is replacing before committing.

Source

bilibili/Index-Translate on GitHub.