Index-Translate

Index-Translate is Bilibili's family of multilingual translation models built on Qwen3.5. The text models cover 150 languages and follow translation instructions around terminology, formatting, and content preservation. The model family also extends into speech-to-speech dubbing, subtitle translation, syllable-controlled translation, and long-document translation. Repo, Hugging Face checkpoints, a technical report, and an online demo are all public.

Key Features

• 150-language text translation: Index-Translate text checkpoints handle a broad language inventory and follow translation instructions.
• Index-Echo: Produces translated subtitles or speech conditioned on the source speaker's voice.
• Index-Homura: Adjusts translations toward a specified target syllable count.
• Index-NativeLong: Translates complete documents with context across passages.
• Multiple sizes: 2B, 9B, and 35B-A3B preview checkpoints are listed for text models.
• Open release: GitHub, Hugging Face, ModelScope, and an online demo are all public.

Use Cases

Who Should Use This Tool?

• Localization teams: Translate documents, subtitles, and structured content while preserving terms and formatting.
• Video workflow builders: Index-Echo packages make subtitle and dubbing pipelines possible from a local model.
• Researchers: The technical report and benchmark data give transparency into training and evaluation choices.

Problems It Solves

1. Translation beyond plain text: Structured JSON, memes, community phrasing, speech, and long documents are handled by different members of the family.
2. Instruction control: You can ask for terminology, style, context, and output-length constraints rather than relying on a freeform chat model.
3. Local and private pipelines: Open checkpoints on Hugging Face and ModelScope let you run translation on your own infrastructure.

Pricing

Advantages & Unique Selling Points

Compared to Competitors: The family is unusually broad, mixing text, speech, syllable constraints, and long documents in one release. The 9B model's card cites strong scores on its own technical report, including FLORES COMET-22 0.8789, WMT26 Judge 75.35, and instTrans IFscore 0.8209.

What Makes It Stand Out: Bilibili publishes a full technical report and a public demo, giving users both reproducible research and an interactive way to see examples.

Getting Started

1. Clone the Index-Translate repository.
2. Install vLLM with Qwen3.5 support.
3. Run vllm serve IndexTeam/Index-Translate-2B --host 127.0.0.1 --port 8000.
4. Run the translation example script with --target en and pass your source text.

Integration

• vLLM with Qwen3.5 support for serving.
• OpenAI-style local HTTP API.
• Hugging Face and ModelScope for model distribution.
• Dedicated inference guides for Echo, Homura, and NativeLong packages.

Frequently Asked Questions

Does it support 150 languages?

The text models do. Speech and long-document packages list separate language coverage.

Is it usable for subtitles?

Index-Echo S2TT translates speech to subtitles with packaged scripts for zh-en, zh-ja, and zh-es directions.

Does it do dubbing?

Index-Echo S2ST performs speech-to-speech translation. Set the source inference defaults carefully because the recommended input window is around 30 seconds per utterance.

Is it Apache-2.0?

The model cards list Apache-2.0.

Alternatives

If Index-Translate is not the right fit, consider these alternatives:

• Qwen3.8-Max: A general multimodal model when you need translation as part of broader agent work.
• MiniMax H3: Useful if you need a production video/audio-capable model from a major provider.
• Deepgram Nova-2: A hosted ASR option when transcription accuracy matters more than open translation weights.

Tips & Best Practices

1. Start with the 2B text model to validate output, then scale to 9B or 35B as quality demands.
2. Use Index-Echo S2ST only with short utterance chunks to avoid budget limits.
3. Verify the exact model IDs in the live repo; Index-NativeLong is released under Index-Nailong model names.

Conclusion

Index-Translate is a serious open multilingual translation family with capable text, speech, and long-document variants. It is most valuable to teams that need local, controllable translation pipelines and care about instruction following rather than only raw MT scores. Like every new model family, benchmark claims should be validated on your own content.