Bilibili’s Index LLM team released the Index-Translate model family on September 30, 2026, putting a multilingual translation project from one of China’s major internet companies into the open-source ecosystem. The release centers on text models that cover 150 languages, including Chinese and English, while also introducing derivatives for speech, dubbing, target-syllable control, and long documents. PanDaily independently reported the same release date.
The publication matters because Index-Translate is presented as a family rather than a single checkpoint. Bilibili’s official materials identify Index-Translate-2B and Index-Translate-9B as released text checkpoints. They also identify Index-Translate-35B-A3B (preview), while making clear that the 35B version is currently in training and that the repository describes the official 35B-A3B release as planned. That distinction is central to reading the announcement accurately: the family includes a preview-stage larger model, not a final official 35B model.
The project and its released model materials use the Apache-2.0 license, according to Bilibili’s official repository and README. That licensing choice makes the release part of a broader open-model pattern already visible in China’s AI sector, where the terms attached to model weights and materials can shape who is able to test, adapt, and deploy them. EastFrontier previously examined that pattern in Tencent’s open-source Hy-MT2 translation models, a relevant comparison because both releases position translation models as downloadable technical resources rather than closed APIs alone.
Index-Translate’s 150-Language Text Coverage Sets the Core Scope
The 150-language figure applies to the Index-Translate text models. It should not be extended to every derivative in the family. Bilibili’s official project materials give the speech and long-document derivatives narrower, separately stated language coverage, so the headline figure belongs specifically to the text checkpoints.
That separation also clarifies what Bilibili has released. Index-Echo is the family’s name for speech translation and dubbing. Index-Homura is the derivative for target-syllable-controlled translation. Index-NativeLong addresses long-document translation, with the model IDs IndexTeam/Index-Nailong-2B and IndexTeam/Index-Nailong-9B published in the official materials. These names describe distinct translation tasks within the same project, rather than proof that one model handles every modality or document setting at the same scale.
For developers, the structure offers several entry points. The 2B and 9B text checkpoints provide the released size labels, while the preview status of Index-Translate-35B-A3B signals a larger version that is still moving through training and release planning. The available evidence does not establish adoption, sales, user counts, or financial performance for Bilibili, and the release itself should be evaluated on the materials that are available rather than on unsupported claims about commercial reach.
Bilibili’s project also fits into EastFrontier’s earlier coverage of China’s open-model push. The common thread is not a claim that every open model has the same capabilities, but that licensing and public release materials make model families available for outside examination. In Index-Translate’s case, the family’s scope is defined by language coverage, task-specific derivatives, and a stated Apache-2.0 license.
Bilibili Reports Tiered Benchmark Results Across Three Checkpoints
Bilibili’s official post-training comparison reports results for the preview 35B-A3B model and the released 9B and 2B checkpoints across five named evaluations. On FLORES COMET-22, the reported scores are 0.8794 for Index-Translate-35B-A3B (preview), 0.8789 for the 9B checkpoint, and 0.8655 for the 2B checkpoint. On the WMT26 judge score, the same three models record 76.76, 75.35, and 60.26, respectively.
The comparison continues with instTrans IFscore, where the reported values are 0.8336, 0.8209, and 0.7569. On MEME, Bilibili reports 0.7405 for the preview model, 0.7387 for 9B, and 0.6443 for 2B. On low-resource FLORES COMET-22, the values are 0.8168, 0.7992, and 0.7377.
Those figures are company- and team-reported benchmark results. They are not independent validation, and they do not establish an unconditional global ranking. The useful information in the table is narrower: Bilibili’s own comparison places the preview 35B-A3B model above the 9B and 2B checkpoints on each named metric, while the 9B model is ahead of 2B on the same comparison set. Readers should keep that attribution attached to every performance claim because the available evidence does not provide outside testing.
The benchmark list also shows how the project is being presented technically. FLORES COMET-22 appears in both general and low-resource forms, while WMT26 judge score, instTrans IFscore, and MEME add separate evaluation lenses. The release therefore supplies a multi-metric comparison, but not a basis for converting those scores into claims about real-world deployment, language-by-language leadership, or superiority over unnamed systems.
Planned Evaluations Keep the Open-Source Release Incomplete
Bilibili’s repository says that planned benchmark releases, including instTrans, SandGlass-V2, nailong-bench, and meme-bench, are not yet open-sourced. The technical report describes the current evaluations, but the associated planned benchmark materials remain unavailable according to the official repository. That creates a clear boundary around what outside users can assess from the release today.
The same boundary applies to the model family’s largest label. Index-Translate-35B-A3B (preview) is currently in training on the official site, and the repository says the official 35B-A3B version is planned. Calling the preview a final model would overstate the status conveyed by Bilibili’s own materials. The released 2B and 9B checkpoints are the firmest basis for discussing what is available now, alongside the separately named derivatives.
The release’s next stage will therefore be defined by what Bilibili publishes, not only by the initial announcement. A final 35B-A3B checkpoint, broader benchmark availability, and clearer external testing would each add information beyond the current fact set. Until then, Index-Translate is best understood as an Apache-2.0 multilingual text-model release with a wider derivative family, a preview-stage larger checkpoint, and performance numbers that remain explicitly Bilibili-reported.
For China’s open-source AI field, the immediate signal is the packaging of translation capabilities into a public family with named checkpoints and specialized variants. For engineers, the practical questions are which language and task each derivative supports, which model ID is appropriate, and what the available evaluation materials actually measure. Bilibili has supplied the initial map. The completeness of that map will depend on the planned models and benchmark releases that have not yet been opened.