Index-Translate: Bilibili's open translators for 150 languages
Bilibili's Index team released Index-Translate, Apache 2.0 translation models in 2B, 9B and 35B-A3B sizes for 150 languages, with image translation and glossary, format and tone controls.
The Index LLM team at Bilibili, a Chinese video platform, published Index-Translate on Hugging Face in late September 2026, with a technical report dated 30 September. It is a family of translation models built on Qwen3.5 in 2B, 9B and 35B-A3B sizes, all under Apache 2.0. The 35B-A3B model and the GGUF files followed in early October, and the 35B-A3B is still marked a preview.
What is new
The models cover 150 languages and take text or images. The main idea is instruction following in translation. Prompts can carry hard constraints, such as an enforced glossary, JSON, CSV or Markdown structure, code blocks and placeholders, and soft constraints, such as tone or the domain meaning of an ambiguous term. Image translation runs through the model’s own vision encoder with no separate OCR step, although the report notes that its own training used no image examples and that quality varies by size.
Training began with a shared multilingual stage of 167.77 billion tokens. Specialists for general translation, instruction following and meme translation followed, trained with supervised fine-tuning and reinforcement learning, merged by interpolation and finished with multi-teacher on-policy distillation. In the report’s main table the 9B model scores 0.8789 on FLORES with COMET-22 and 75.35 on a WMT26 judge scale. It is ahead of Hy-MT2 7B and TranslateGemma 12B on eight of nine columns, with the widest gaps on instruction following and its own meme test. These are the team’s own tests, using its own instTrans and MEME benchmarks and GPT-5.6-Sol as judge.
Running it
The 2B weights are 4.6 GB in BF16 and the 9B weights 19.3 GB, and the GitHub guide gives 24 GB of GPU memory as a starting point for the 9B. The 35B-A3B preview has about 36 billion parameters, with 3 billion active, and is 71.9 GB. GGUF files for llama.cpp are 1.31 GB for the 2B and 5.78 GB for the 9B at Q4_K_M, and separate mmproj files add image translation.
Serving with vLLM needs a build with Qwen3.5 support. The card’s prompt templates are written in Chinese, its example client decodes greedily with thinking switched off, and the shipped configuration allows 262,144 tokens of context while the examples serve 32,768.
License
Every model card and the GitHub repository declare Apache 2.0, with no revenue cap, region limit or naming rule, and Qwen3.5 is Apache 2.0 too. On 5 November the GitHub README added Index-Translate-Omni checkpoints that bring Chinese and English speech recognition and speech-to-text translation to the same family. Separate Index-Echo, Index-Homura and Index-NativeLong models, which cover speech, dubbing and full documents, are outside our entry.
Why it matters
Index-Translate lists the most languages of the open translators we track, 150 against 33 to 55 for the others, and it packs glossary, format and tone controls and image input into one Apache 2.0 family.
Every score comes from the team’s own report, the prompts are in Chinese, and the card says combined constraints and low-resource directions remain harder. We found no independent comparison. Our Index-Translate vs TranslateGemma page sets out the trade-offs, and the translation models list shows the field.
Sources
Models in this story
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Index-Translate Bilibili | 2.27B – 36B 3B active | 256K | Apache 2.0 (Permissive) | 1.44 GB Laptop or phone | |
| Hy-MT2 Tencent | 2.04B – 30.1B 3B active | 256K | Apache 2.0 (Permissive) | 1.41 GB Laptop or phone | |
| TranslateGemma Google DeepMind | 4.3B – 27.4B | 2K | Gemma Terms (Conditional) | 2.97 GB Laptop or phone | |
| Qwen3.5 Alibaba Qwen | 873M – 403B 17B active | 256K | Apache 2.0 (Permissive) | 0.602 GB Laptop or phone |