Bilibili’s Index-Translate Ships 150-Language Weights With a 35B MoE Preview

Bilibili’s Index LLM Team released Index-Translate on September 30, 2026—an Apache-2.0 multilingual translation family built on Qwen3.5, with 2B and 9B dense text checkpoints plus a 35B-total / 3B-active MoE preview covering 150 languages, and sibling packages for speech translation, syllable-controlled dubbing, and full-document work. Weights are on Hugging Face under IndexTeam and mirrored on ModelScope; inference clients, a technical report, and an online demo ship from github.com/bilibili/Index-Translate. The repository’s news line dates the text-model drop to 2026-09-30.

The text models are the spine. They handle ordinary sentence and article translation, structured payloads such as JSON with preserved keys and markup, and instruction constraints for terminology, formatting, style, structure, context, and length. The 35B-A3B preview is labeled as such on the Hub card: all of its published scores refer to the preview evaluated in the report, and the project TODO still lists an official non-preview 35B release. Serving uses a vLLM build with Qwen3.5 MoE support; the card’s configuration sets max_position_embeddings=262144, while the public serving examples recommend --max-model-len 32768 for practical GPU memory. Decoding defaults in the shipped translate.py client are greedy at temperature 0 with thinking disabled through chat_template_kwargs.

Benchmark tables on the 35B-A3B preview card put FLORES COMET-22 at 0.8794 and instTrans IFscore at 0.8336—highest among the systems listed in that comparison for those two columns—alongside 0.6901 instTrans quality, 76.76 WMT26 Judge, and 0.7405 MEME. On low-resource FLORES minor pairs the same preview reports 0.8168 COMET-22 / 0.7164 XCOMET-XXL with a 2.4% off-target rate. Domain COMET-22 figures include 0.8962 on MuST-Cinema subtitles and 0.8681 on biomedical text. General-capability checks after translation training land at 77.6% C-Eval and 49.1% GPQA-Diamond on the card’s matched subsets. Those are report-protocol numbers, not a claim that every language pair or constraint stack will match the headline axes.

Training is shared mid-training then specialist post-training. The card describes 167.77B tokens of multilingual mid-training with staged mixes of general, parallel, and monolingual (and later pivot) data, followed by specialist SFT and RL for general translation, instruction following, and meme translation; expert interpolation and multi-teacher on-policy distillation (MOPD) combine complementary specialists. Interpolation weights of 0.8 / 0.1 / 0.1 are stated for the evaluated 2B and 9B models and are explicitly not assigned to the 35B-A3B preview.

The family extends beyond plain text. Index-Echo ships speech-to-text (S2TT) and speech-to-speech (S2ST) packages at 2B and 9B with narrower packaged language pairs than the text models’ 150-language inventory. Index-Homura targets a requested syllable count for dubbing-style length control. Index-NativeLong—published under Hub IDs IndexTeam/Index-Nailong-2B and Index-Nailong-9B—uses fixed templates for full-document translation with long context presets (the serving table lists 262144 for NativeLong-2B). The repository also includes a browser extension path and a video-dubbing pipeline that extract audio, translate, and realign to video.

The repository ships a shared OpenAI-compatible client against a local vLLM endpoint, language-code mapping into Chinese prompt templates aligned with training, and separate entry points for Homura syllable budgets and NativeLong document directions. Echo S2TT and S2ST use dedicated scripts and the packaged language pairs on the model table. The demo at index-translate.bilibili.com shows terminology preservation, meme-sense disambiguation, syllable targets, and document-level name consistency beside comparison systems. Planned follow-ons in the TODO list include an official non-preview 35B-A3B, open-sourcing of instTrans / SandGlass-V2 / nailong-bench / meme-bench, broader Echo language coverage, and larger models—so today’s drop is a family launch with an unfinished checklist, not a claim that every benchmark suite is already public.

instTrans separates translation quality from instruction adherence across ten constraint types on Chinese-to-20-language tasks; MEME scores community expressions across thousands of Chinese-to-English cases with hundreds of senses. Low-resource minor-pair and minor-instruction splits report off-target rates so teams can see whether a model answers in the wrong language rather than only whether COMET looks high. Closed API baselines such as DeepSeek-V4.1-Flash, GPT-5.6-Sol, and Gemini 3.5 Flash Lite appear in the same tables; the preview leads some columns and trails others—useful for procurement rather than a single “best translator” slogan.

For Writeble’s opensource desk, Index-Translate is a September 30 open-weights translation stack with a clear Apache-2.0 license, a published MoE preview, and speech and document siblings—not a single Hub dump without serving code.