huggingface.co via Reddit

Bilibili Open-Sources Index-Translate 35B MoE for 150 Languages

By Published Updated
5 sources tracking this story
Open Source China AI Multimodal ai-business

TL;DR

  • The 35B-A3B MoE activates only 3B parameters per token, keeping inference cost near the 9B model on consumer hardware.
  • Index-Translate-9B matches or exceeds GPT-5.6-Sol and Gemini 3.5 Flash Lite on both WMT26 judge scores and FLORES COMET-22 benchmarks.
  • Index-Homura adds syllable-constrained dubbing translation via reinforcement learning, a capability absent from existing open-source translation stacks.

Bilibili's Index LLM Team has put out an Apache 2.0 translation model family whose top preview is a 35-billion-parameter mixture-of-experts activating 3 billion at a time, covering 150 languages. The model card on Hugging Face lists Index-Translate-35B-A3B-preview as built on a Qwen3.5 backbone, alongside dense 2B and 9B siblings for text and separate 2B/9B checkpoints for the Echo (speech-to-text and speech-to-speech translation), Homura (syllable-controlled dubbing) and NativeLong (full-document) tasks.

The reported numbers are the pitch. On FLORES the preview lands a 0.8794 COMET-22, and on low-resource pairs the card puts its off-target rate at 2.4%, the lowest in its comparison set. Domain scores reach 0.8962 COMET-22 on MuST-Cinema subtitles and 0.8681 on biomedical text. The general-capability sheet shows 77.6% on C-Eval and 49.1% on GPQA-Diamond.

Training used 167.77 billion tokens of multilingual mid-training, split into a constant stage mixing general, parallel and monolingual data at 1:1:1 and a decay stage weighted 1:4:2 toward pivot data. Post-training runs through specialist SFT plus RL for general translation, instruction following and meme translation; a general-translation RL step judged by XCOMET-XXL with language-validity and adequacy checks; an instruction step the card describes as "Rubric-as-Reward with hard checks and graded constraints"; parameter interpolation of the specialists; and a final multi-teacher on-policy distillation.

The card qualifies its own claims: "Scores vary across languages and tasks. Combined constraints and low-resource directions remain harder than core-language translation, and successful individual examples do not guarantee that every instruction will be followed."

Serving is a `vllm serve` one-liner with a shipped 262,144-token context window and a recommended `--max-model-len 32768`. This is the second Chinese open-weights release we've logged this week, after InclusionAI's 560B Ling-3.1-flash.

What others are reporting

Coverage cluster as of 24h after publish

  1. GitHub (Bilibili official) Read →

    First-party repository with full benchmark tables comparing all four model branches against GPT-5.6-Sol, Gemini 3.5 Flash Lite, and DeepSeek, plus architecture docs for each branch.

  2. Pandaily Read →

    English-language China tech coverage situating the release in Bilibili's broader open-source AI track record and the Qwen3.5 base choice.

  3. PANews Read →

    Brief item citing IT Home; confirms dubbing and long-document specialized branches ship alongside the base text models at launch.

    The models cover 150 languages, support translation instructions such as terminology and formatting, and extend to subtitle dubbing
  4. Phemex News Read →

    Confirms dual distribution across Hugging Face and ModelScope and that all three text-model sizes are available at release.

    Model weights for 2B, 9B, and 35B-A3B text variants are now openly available on Hugging Face and ModelScope.

Originally reported by huggingface.co

Read the original article →

Original headline: Bilibili Open-Sources Index-Translate Covering 150 Languages, Tops Out at 35B MoE Under Apache 2.0