Index-Translate

Bilibili's Apache 2.0 translation models in 2B, 9B and 35B-A3B sizes for 150 languages, with image translation and glossary control.

In short. Index-Translate is a family of AI translation models from the Chinese video platform Bilibili. You pick a size: a 2B model for a laptop, a 9B model for a single GPU and a 35B-A3B preview for a server. They translate between 150 languages, read text inside images, and follow instructions such as keeping a glossary term, a JSON structure or a polite tone. The license is Apache 2.0.

Summary

Bilibili's Index LLM team published Index-Translate on Hugging Face in late September 2026, together with a technical report dated 30 September, and added the 35B-A3B preview and the GGUF files in early October. It is built on Qwen3.5 and comes in three sizes: a 2B model of about 2.27 billion parameters, a 9B model of about 9.65 billion and a 35B-A3B mixture-of-experts model of about 36 billion parameters with 3 billion active, which is still labeled a preview. All of them cover 150 languages and take text or images, and the cards say image translation is done by the model itself through a companion vision encoder instead of an external OCR tool. The report notes that Index-Translate's own training used no image examples and that image results vary by model size. The shipped configuration allows 262,144 tokens of context, and the examples serve it with 32,768.

The team's main idea is instruction following in translation. The prompts support hard constraints, such as an enforced glossary, JSON, CSV and Markdown structure, code blocks and placeholders, and soft constraints, such as tone and the meaning of an ambiguous term in its domain. Training began with a shared multilingual stage of 167.77 billion tokens. Specialist models for general translation, instruction following and meme translation were then trained with supervised fine-tuning and reinforcement learning, merged by interpolation with weights of 0.8, 0.1 and 0.1 for the 2B and 9B models, and finished with multi-teacher on-policy distillation. The report says the 9B model scores 0.8789 on FLORES with COMET-22 and 75.35 on a WMT26 judge scale, and that it has the highest score and the lowest off-target rate on the report's low-resource instruction-following test among the systems compared. In the main table it is ahead of both Hy-MT2-7B and TranslateGemma-12B on eight of nine columns, by the widest margin on instruction following, and behind both on one reference-based score of the IFMTBench benchmark. These are the developer's tests, and instTrans and MEME are its own benchmarks, with their release only tracked as a plan on GitHub.

Plan on about 19.3 GB for the 9B weights in BF16 and 4.6 GB for the 2B, and the GitHub guide gives a starting point of 24 GB of GPU memory for the 9B. GGUF files for llama.cpp are 1.31 GB for the 2B model 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, and the card gives prompts written in Chinese that match the training format, with thinking switched off and greedy decoding. The GitHub repository also lists Index-Translate-Omni checkpoints that add Chinese-English speech recognition and speech-to-text translation, and separate Index-Echo, Index-Homura and Index-NativeLong models for speech, dubbing and full documents. The license is Apache 2.0. On 11 November 2026 Hugging Face showed about 92,000 downloads of the 9B GGUF repository since release, far more than the 3,900 of the 9B weights.

Good for

  • Self-hosted translation across 150 languages, including many low-resource languages, on one GPU with the 9B model or on a laptop with the 2B GGUF
  • Translation that must follow a glossary, keep JSON, CSV, Markdown, code or placeholders unchanged, or match a tone and genre
  • Translating text inside screenshots, posters and photos without a separate OCR step
  • A commercially unrestricted translation model, since the license is Apache 2.0

Watch out for

  • The quality claims come from the team's own report and its instTrans and MEME benchmarks, and the report uses GPT-5.6-Sol as a judge, so test on your own text
  • The 35B-A3B model is a preview, and the model card says combined constraints and low-resource directions remain harder than core languages
  • The recommended prompts are written in Chinese, even when the source and target languages are others
  • Serving with vLLM needs a build with Qwen3.5 support, image translation in GGUF needs the separate mmproj file, and image quality varies by model size
  • Speech, dubbing and whole-document translation belong to other models in the family, not to this entry

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the Index-Translate-2B (GGUF Q4_K_M) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
Index-Translate-2B BF16 2.27B 5.22 GB 2.61 GB 1.57 GB Laptop or phone
Index-Translate-2B (GGUF Q4_K_M) INT4 2.27B 5.22 GB 2.61 GB 1.44 GB Laptop or phone
Index-Translate-9B BF16 9.65B 22 GB 11 GB 6.66 GB Laptop or phone
Index-Translate-9B (GGUF Q4_K_M) INT4 9.65B 22 GB 11 GB 6.36 GB Laptop or phone
Index-Translate-35B-A3B (preview, MoE) BF16 36B 3B active 83 GB 41 GB 25 GB Gaming GPU

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is Index-Translate?
Index-Translate is a family of AI translation models from the Chinese video platform Bilibili. You pick a size: a 2B model for a laptop, a 9B model for a single GPU and a 35B-A3B preview for a server. They translate between 150 languages, read text inside images, and follow instructions such as keeping a glossary term, a JSON structure or a polite tone. The license is Apache 2.0.
Can I use Index-Translate commercially?
Yes. Index-Translate is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Index-Translate open source?
Index-Translate is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Index-Translate on my own computer?
Yes. The smallest variant, Index-Translate-2B (GGUF Q4_K_M), needs about 1.44 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was Index-Translate released?
Index-Translate was released by Bilibili on September 2026.
How long a context window does Index-Translate have?
Index-Translate supports a context window of up to 256K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
TranslateGemma Google DeepMind 4.3B – 27.4B 2K Gemma Terms (Conditional) 2.97 GB Laptop or phone
Hy-MT2 Tencent 2.04B – 30.1B 3B active 256K Apache 2.0 (Permissive) 1.41 GB Laptop or phone
PLaMo 3 Translate Preferred Networks 890M – 31.8B 32K PLaMo Community (Conditional) 0.614 GB Laptop or phone
Qwen3.5 Alibaba Qwen 873M – 403B 17B active 256K Apache 2.0 (Permissive) 0.602 GB Laptop or phone