Bilibili Open-Sources 150-Language Translation Models
A permissively licensed model family brings 150-language translation and specialized localization variants to builders.
30-Second TL;DR
What Changed
Index-Translate supports 150 languages.
Why It Matters
Apache-2.0 licensing could make multilingual translation more accessible for commercial and research applications. The specialized variants may be useful for localization pipelines that require timing, pronunciation or long-context handling.
What To Do Next
Download the Apache-2.0 2B and 9B checkpoints and benchmark them on your target language pairs and document lengths.
Key Points
- •Index-Translate supports 150 languages.
- •Preview text models are available in 2B, 9B and 35B-A3B configurations.
- •Additional variants target dubbing, syllable control and long documents.
Deep Insight
Background and context from public sources — not the original article. 6 sources cited.
Enhanced Key Takeaways
- •Index-Translate models are built upon Alibaba's Qwen3.5 base architecture and hosted openly on Hugging Face and ModelScope.
- •The 35B-A3B variant employs a Mixture-of-Experts (MoE) configuration that activates only 3 billion parameters per token out of 35 billion total parameters.
- •A dedicated speech-to-speech branch named Index-Echo enables voice-cloning dubbing across Chinese, English, Spanish, and Japanese while retaining speaker characteristics.
- •The specialized Index-Homura model introduces syllable-budget enforcement for video subtitle syncing, reaching 81.92% compliance within a 10% syllable margin on the SandGlass benchmark.
- •Bilibili integrated domain-specific fine-tuning for ACG community slang, gaming terminology (such as Final Fantasy XIV jargon), and internet memes alongside a local browser extension for edge deployment.
Technical Deep Dive
- Foundational Architecture: Built on Alibaba's Qwen3.5 architecture, available in dense (2B, 9B) and sparse MoE (35B-A3B, 35B total with 3B active parameters per token) configurations.
- Syllable-Control Variant (Index-Homura): Specifically calibrated to adhere to target syllable constraints for video synchronization; evaluated on the SandGlass benchmark with an 81.92% compliance rate within a 10% syllable tolerance.
- Speech and Dubbing (Index-Echo): Provides speech-to-speech translation with speaker voice preservation and vocal characteristic matching across bidirectional Chinese, English, Spanish, and Japanese pairs.
- Long-Context Modeling (Index-NativeLong / Index-Nailong): Tailored for multi-thousand token contexts to maintain cross-passage consistency and reduce named entity translation drift.
- Formatting and Robustness: Retains structured output tags, Markdown, and JSON schema during translation while mapping colloquialisms evaluated via the instTrans and MEME benchmarks.
Future ImplicationsAI analysis grounded in cited sources
Timeline
- 2026-10Bilibili open-sources the Index-Translate model suite under the Apache-2.0 license
Sources (6)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
Weekly AI Recap
Read this week's curated digest of top AI events →
Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: Pandaily ↗
This is a summary, not the original. Read the source, or get the weekly briefing.
The weekly digest
One email a week. Unsubscribe anytime.