Pandaily (@thePandaily) on X

What it claims

Bilibili's Index-Translate (released September 30) is an instruction-following translation model family built on Qwen3.5, covering 150 languages. The models follow constraints on terminology, formatting and content that must stay untranslated. On the team's own tables, the 9B is close to the 35B-A3B preview on FLORES. It leads every compared model on low-resource instruction following. It trails some large general models on WMT26 judge scores.

Evidence

All figures are from Bilibili's own benchmarks as reported in this news post:

• Low-resource instruction following: the 9B posted the best instruction score (0.7725) and the lowest off-target rate (3.47%) among all compared models. The post does not list the other models or their scores.
• Syllable control: Index-Homura-9B lands within 10% of the requested syllable count 81.92% of the time on the in-house SandGlass test.
• Long-document consistency: the test was a fantasy text of about 32,000 tokens where a character's name could also be read as a royal title.
• The long-document model (Index-NativeLong) kept the name consistent throughout.
• The base 9B, translating chunk by chunk, drifted to other renderings.
• This is a single qualitative example, not a metric.
• Format preservation: the 9B translated a JSON game-maintenance notice into Korean. It kept the JSON structure, the star symbols and a hashtag the user asked to preserve.

The practical reading: if your workload is general high-resource translation, a large general model may score higher on WMT26. The case for Index-Translate is constrained or structured translation, low-resource languages, and self-hosting at 2B or 9B.

How to use it

• Sizes: 2B, 9B and 35B-A3B. The 35B-A3B is a preview.
• Licence and hosting: Apache-2.0, on Hugging Face and ModelScope. A technical report, GitHub code and an online demo accompany the release.
• Long-document model: published under model IDs named Index-Nailong, not "NativeLong". Search for that name.
• Index-Echo coverage: the packaged speech-to-speech release covers only:
• Chinese to English, Spanish and Japanese
• English to Chinese, Spanish and Japanese

The 150-language coverage applies to the text models.

• Extras in the repository: a browser extension that translates web pages with a locally deployed model, and a video dubbing pipeline.
• Commands: the page gives no install or run commands.

Limits

• Self-reported results: every number comes from the team's own benchmarks, relayed by a news outlet rather than read from the technical report.
• Benchmarks not yet public: the instruction-following, community-slang and SandGlass benchmarks are not yet open-sourced. Releasing them is listed as a plan, so you cannot reproduce those scores today.
• 35B-A3B is a preview: an official release is still pending.
• Index-Echo is narrow: it covers six directions. Adding languages is a stated future plan.
• Domain bias: the examples skew toward Bilibili's own community content, such as gamer slang and fantasy fiction. Check results on your own domain before relying on the slang-handling claims.