Bilibili Index-Translate-35B-A3B Open Multilingual Translation Model
Bilibili's Index team published Index-Translate-35B-A3B-preview on 2 October 2026, describing it as the 35B-total, 3B-active mixture-of-experts (MoE) multilingual text translation model of the Index-Translate family, built on Qwen3.5.
Tool Overview
Features, steps and FAQ below
Features
- ✓ Bilibili says the model supports text translation across 150 languages and follows translation instructions involving terminology, formatting, style, structure, context and output length
- ✓ Bilibili says it is a 35B-total, 3B-active MoE model that shares a common multilingual mid-training recipe with its 2B and 9B siblings, followed by task-specific post-training
- ✓ Bilibili says the weights are open under an Apache 2.0 license on Hugging Face and ModelScope, shipped with an online demo, GitHub code, a technical report, browser-extension code and a video-dubbing pipeline
- ✓ The model card states the shipped configuration has a 262,144 context window while the card examples serve 32,768; defaults are greedy decoding (temperature=0), a max_tokens of 1024 and thinking disabled
- ✓ Bilibili says Index-Translate is a family covering text, speech, controlled dubbing and long-document translation, alongside Index-Echo S2TT/S2ST, Index-Homura (target-syllable translation) and Index-NativeLong (full-document translation)
How to Use
- Open the Hugging Face model card (huggingface.co/IndexTeam/Index-Translate-35B-A3B-preview) or the ModelScope page for details and license
- Clone the GitHub repository (github.com/bilibili/Index-Translate) and prepare your inference environment
- Run the official translate.py inference client and use --source and --target to set the input and output languages (provide the full language name for unmapped codes)
- You can also try the online demo (index-translate.bilibili.com) or follow the prompt guide to set terminology, structure, style and context instructions
FAQ
What is Index-Translate-35B-A3B-preview?
A 35B-total, 3B-active MoE multilingual text translation model published by Bilibili's Index team on 2 October 2026, which Bilibili says is built on Qwen3.5 and supports 150 languages.
How many languages can it translate?
The model card says it supports text translation across 150 languages and follows translation instructions covering terminology, formatting, style, structure, context and output length.
Is the model open source?
Bilibili says the weights are open under an Apache 2.0 license on Hugging Face and ModelScope, with GitHub code, a technical report and an online demo.
What is the context length?
The model card states the shipped configuration uses a 262,144 context window and the card examples serve 32,768; exact protocols are in the technical report.
How does it differ from the 2B and 9B versions?
Bilibili says all three belong to the Index-Translate family and share the multilingual mid-training recipe; the 35B-A3B is the 35B-total, 3B-active MoE version and is stronger at following complex translation instructions.