Index-Translate: Bilibili releases an open translation model family for 150 languages
Bilibili's Index team has released Index-Translate, a family of translation models built on Qwen3.5, under the Apache 2.0 license. The text models cover 150 languages, including Turkish and Arabic, and the family also includes models for speech translation, syllable-controlled dubbing and long docum
The Index team at Bilibili, one of China's largest video platforms, has released Index-Translate, a family of translation models. According to the team's GitHub page, the technical report, an online demo and the text models were announced on September 30, 2026; the models were uploaded to Hugging Face by October 2, and compressed builds that need less memory followed on October 3. The open weights are available for anyone to download, and the whole family is released under the Apache 2.0 license, so it can be used commercially and run on your own servers or on a local computer.
The family is built on Alibaba's Qwen3.5 models and covers four jobs: text translation, speech translation, translation tuned to a syllable count for dubbing, and long-document translation. According to the model card, the text models cover 150 languages. In the technical report's language list, Turkish and Arabic are among the 22 core languages.
Text translation models in three sizes
On the text side there are two models with 2 billion and 9 billion parameters, plus Index-Translate-35B-A3B-preview. The latter is a mixture-of-experts (MoE) model: it has 35 billion parameters in total, but only about 3 billion are active at each step. This usually makes it faster than a dense model of the same size, although all of its files still need to fit in memory. Beyond plain translation, the models follow instructions such as sticking to a supplied glossary, preserving JSON, CSV and Markdown formatting, code blocks and placeholders like {variable}, and adapting the text to a formal or casual tone. The team says it also aims to translate slang, community nicknames and jokes by their intended meaning when a literal translation would fail.
Speech, dubbing and long-document models
Index-Echo translates speech and comes in two forms. The S2TT model turns an audio or video file into bilingual subtitles with sentence timestamps; the packaged release translates from Chinese into English, Japanese and Spanish. The S2ST model re-voices the translated speech, using the source speaker's voice as a reference, and supports six directions: Chinese into English, Spanish and Japanese, and English into Chinese, Spanish and Japanese. Adding more languages to Echo is on the team's to-do list.
Index-Homura prepares lines for dubbing and timed subtitles by steering the translation toward a requested syllable count, and it can be combined with a glossary. Index-NativeLong translates a document in one pass without splitting it into chunks, which helps keep names and terms consistent across a book or a script. The released configuration of the 9B version supports a context of up to 229,376 tokens. The packaged templates currently cover Chinese to and from English and Chinese to and from Japanese. On Hugging Face its IDs are Index-Nailong-2B and Index-Nailong-9B. The speech and dubbing models also come in 2B and 9B sizes.
Results in the team's report
According to the technical report, the 35B-A3B preview leads the systems compared in the report with a COMET-22 score of 0.8794 on FLORES and an instruction-adherence score of 0.8336 on the instTrans instruction translation test. The comparison includes models such as GPT-5.6-Sol, Gemini 3.5 Flash Lite, DeepSeek-V4.1-Flash and TranslateGemma; on the WMT26 evaluation, GPT-5.6-Sol, DeepSeek-V4.1-Flash and Gemini 3.5 Flash Lite score higher. In low-resource languages, 2.4% of the 35B-A3B model's outputs come out in the wrong language. On the team's SandGlass test, Index-Homura-9B lands within 10% of the requested syllable count 81.92% of the time. The team also reports that the Index-Echo models place subtitle start times with less drift than the other models compared.
Running it locally and other ways to try it
The model cards show how to serve the text models through vLLM as an OpenAI-compatible server. According to the GitHub guide, in unquantized BF16 the 2B model needs about 8 GB of GPU memory and the 9B model about 24 GB. For ordinary computers, GGUF files for llama.cpp are also available: in the recommended 4-bit build, the 2B model takes about 1.3 GB, the 9B model about 5.8 GB and the 35B-A3B model about 21.7 GB. For servers there are FP8 builds and NVFP4 builds aimed at NVIDIA's Blackwell GPUs. If you want to try it without installing anything, there is a demo at index-translate.bilibili.com and, according to the GitHub page, an OpenAI-compatible API for the 35B-A3B model available since October 4.
The 35B-A3B model is still a preview, and the official release is on the team's GitHub to-do list. According to the model card, tasks that combine several instructions and low-resource language directions remain harder than core-language translation, and individual successful examples do not guarantee that every instruction will be followed. Since long-document translations can have missing or repeated passages, the card recommends reviewing the output. Even though Turkish and Arabic are among the core languages, the safest approach is to test the models on your own texts and language pairs. For a fuller picture, follow evaluations from independent researchers as well.
Sources: Hugging Face: Index-Translate collection, Index-Translate-35B-A3B-preview model card, Index-Translate-9B model card, Index-Echo-S2ST-9B model card, Index-Homura-9B model card, Index-NativeLong-9B model card, Technical report (arXiv), GitHub: bilibili/Index-Translate.
