--- title: "Bilibili Open-Sources Index-Translate, a 150-Language Model Family for Video Dubbing" date: 2026-10-29 category: Generative Media site: NeuroAI canonical: https://neuroai.site/a/na-aigc-bilibili-index-translate language: en --- # Bilibili Open-Sources Index-Translate, a 150-Language Model Family for Video Dubbing > Bilibili (哔哩哔哩) released Index-Translate under Apache-2.0 on 30 September 2026, a Qwen3.5-based family spanning text, subtitle and voice-dubbing translation. A fan in São Paulo finds a forty-minute Chinese game explainer on Bilibili and wants to watch it without reading subtitles. A year ago that meant waiting for a volunteer. Today the platform itself ships the model that can dub it. The interesting part is not the feature — it is that the model is open. ## Key takeaways - Bilibili's (哔哩哔哩) Index LLM team released Index-Translate on 30 September 2026 under the Apache-2.0 license, built on Alibaba's Qwen3.5. - The text models cover 150 languages and ship in 2B, 9B and 35B-A3B (preview) sizes on Hugging Face and ModelScope. - Three specialized branches handle speech subtitles, voice-preserving dubbing, and whole-document consistency. - The 35B-A3B flagship is a Mixture-of-Experts model with 35B total and roughly 3B active parameters and a 262,144-token context. - Benchmark scores are vendor-reported and should be read as claims until independently verified. ## What Bilibili actually shipped On 30 September 2026, Bilibili's (哔哩哔哩) in-house Index LLM team published Index-Translate, a multilingual translation model family built on Alibaba's Qwen3.5 and released under Apache-2.0. The text models cover 150 languages and come in 2B, 9B and 35B-A3B (preview) sizes, with weights on Hugging Face and ModelScope, a technical report, code on GitHub, and an online demo. Pandaily and model-tracking sites such as llm-releases.com independently recorded the same release date, license, and base model, so the core facts are consistent across sources. The flagship 35B-A3B is a Mixture-of-Experts model: 35B parameters in total, about 3B active per token, with a 262,144-token context window. That long context is the point — translation quality collapses when a model forgets what a character was called three pages ago. ## Beyond plain translation Index-Translate is not one model but a family with four jobs. Index-Echo turns speech into translated subtitles (speech-to-text translation) and, in its speech-to-speech form, produces dubbed audio that keeps the source speaker's voice characteristics; the packaged speech-to-speech release covers Chinese to and from English, Spanish and Japanese. Index-Homura adjusts a translation toward a target syllable count, which matters when a line has to fit a scene's runtime. Index-NativeLong translates entire documents and keeps names and terms consistent across paragraphs. The syllable-control branch is the clearest sign the team was thinking about video, not just text. On the team's own SandGlass test, Index-Homura-9B landed within 10 percent of the requested syllable count 81.92 percent of the time. For subtitlers and amateur dubbers, a line that fits the mouth is more useful than a line that is literally correct. The same logic applies to memes and community slang: the project ships a dedicated benchmark for translating internet humor, because a literal rendering of a joke is usually not a translation at all. That focus on register, not just vocabulary, is what separates a community-aware model from a generic machine translator. ## The numbers, and their limits On the team's own benchmarks, the 35B-A3B preview scores 0.8794 on FLORES (COMET-22) and 76.76 on a WMT26 judge score, while the 9B reaches 0.8789 and 75.35. Some larger general models score higher on the same WMT26 table — DeepSeek-V4.1-Flash is listed at 83.55. On low-resource instruction following, the 9B posts the best score among compared models at 0.7725, with the lowest off-target rate at 3.47 percent. Every figure above is vendor-reported; the model-tracking sites explicitly flag them as claims until independently verified, and we have not re-run the evaluations. ## Why a video platform builds translation models Bilibili (哔哩哔哩) is a Chinese video community with a large diaspora and a growing international app. User-generated content lives or dies on whether someone outside the original language can watch it. A platform that owns the translation layer — text, subtitle, and dubbing — controls how far its creators travel. Open-sourcing the models also lets self-hosters build private pipelines and browser extensions, which is how a community translation ecosystem usually forms. ## Honest limitations We rely on Pandaily, llm-releases.com and Chinese AI tool directories; we did not inspect the GitHub repository or reproduce any benchmark. The 150-language figure applies to the text models, and speech and long-document branches list narrower, separately documented coverage, so a team planning a multilingual dubbing pipeline should not assume the full set carries over. The vendor benchmark scores are self-reported and unverified by us. We have not tested output quality, latency, or whether the free demo reflects the released weights. ## What readers can do now - If you localize video, prototype Index-Echo for subtitle generation and Index-Homura for line-length control before paying for a managed dubbing service. - If you run a community platform, study whether open translation weights reduce your dependence on a single closed vendor for cross-language reach. - If you evaluate models, treat the published FLORES and WMT26 numbers as a starting hypothesis and benchmark on your own domain text before trusting them in production. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-aigc-bilibili-index-translate Free to quote with attribution and a link to the original.