Index-Translate: A 2B Translation Model for 150 Languages That Runs on Your Laptop
Most model releases these days are another coding chatbot or another general assistant. Index-Translate, from Bilibili’s Index team, is a refreshing exception: a model family that does one thing — translation — and does it well enough to run the useful version on a laptop.
What it is
Index-Translate is built on Qwen3.5 and covers 150 languages. It is not a single model but a family, each member tuned for a different translation problem:
- Index-Translate — the core text model. Translates prose, structured content, and community/slang expressions, and follows instructions about terminology, formatting, and content preservation.
- Index-Echo — produces translated subtitles or speech conditioned on the source speaker’s voice. The translated audio sounds like the original person.
- Index-Homura — adjusts a translation toward a target syllable count. This is the unglamorous but critical trick behind dubbing that actually matches lip movement.
- Index-NativeLong — translates whole documents while carrying context across passages, instead of mangling consistency one paragraph at a time.
It ships in three sizes: a 35B-A3B mixture-of-experts preview, a 9B, and a 2B. The 2B is the story for local-AI folks.
Why the small one matters
The published benchmarks put Index-Translate on a seven-axis radar — WMT, FLORES, instruction following, low-resource languages, subtitles, internet “MEME” language, and books/fiction — against 14 models including DeepSeek-V4.1-Flash and GPT-5.6-Sol. On translation-specific tasks, the Index models hold their own against those much larger generalists.
That is the whole point, and it is a theme we keep coming back to: a model trained for one job can match a frontier generalist on that job while fitting on hardware you already own. You do not need a 400B-parameter API endpoint to translate a subtitle file. You need a 2B model that was actually trained on translation, running on your GPU, for free, offline.
Where this is genuinely useful
- On-device translation — no API bill, no rate limits, no text leaving your machine. For anything privacy-sensitive (legal, medical, internal docs), that last point is the whole ballgame.
- Dubbing pipelines — Echo (voice-preserving speech) plus Homura (syllable targeting) is a surprisingly complete local toolkit for translated video that lip-syncs and keeps the speaker’s voice.
- Long documents — NativeLong’s cross-passage context is the difference between a translated book that reads coherently and one that renames the protagonist every chapter.
The caveat
Translation quality is domain-dependent, and a 2B model will trail the 9B and 35B versions on the hardest low-resource pairs and the most literary text. The normalized radar scores are not accuracy percentages — they are relative positions, and “competitive with GPT-5.6 on this axis” is not “better at everything.” Benchmark on your language pairs and your content before you wire it into production.
But as a demonstration that the open-model world is maturing past chatbots into specialized, locally-runnable tools, Index-Translate is exactly the kind of release worth paying attention to.