News Weekly
ENहिन्दी
LV 10 XP
0% read
Your progress · 0/5 chapters
About 5 min total
Models & EconomicsISSUE #4 · STORY 39 OF 40Oct 1, 2026COMPANY CLAIM

Bilibili published 150-language translation under Apache 2.0, with the rules in the prompt

Bilibili's Index team released Index-Translate on 1 October 2026 in 2B, 9B and 35B-A3B sizes covering 150 languages, alongside speech, syllable-controlled and long-document variants. Licence terms are clear; every quality and coverage figure is the company's own, with no third-party evaluation found.

A ruled sheet carried through an upright frame on a drafting table and emerging with an identical grid and identical fixed fields, only its free lines rewritten.

Read it your way

CHAPTER 1Picked for Explorer

The 60-second version

Bilibili's Index team open-sourced Index-Translate on 1 October 2026: three sizes built on Qwen3.5, covering 150 languages under Apache 2.0. The distinctive part is not the language count. It is that constraints are written into the instruction.

Three sizes, one familyText models come in 2B and 9B variants plus a 35B-A3B mixture-of-experts preview.
Apache 2.0 throughoutWeights, code and documentation can be used, modified and deployed commercially without copyleft obligations.
Hard and soft constraintsGlossaries, placeholders and JSON or code structure are hard requirements; tone and domain are soft.
Quality is unmeasured so farNo third-party evaluation was found, and performance across 150 languages is likely uneven.
Finish this chapter for +15 XP
A machinist's bench split between a hardwood jig whose part fails to fit its cut-out and a thin spring gauge beside a second part that nothing prevents using.
Flip the switch

From sentence service to constrained pipeline

InterfaceInstructions travel with the textNumbered hard and soft requirements are part of the request, set per job.
TerminologyPinned in the promptA glossary is supplied so the same term renders the same way every time.
FormatPreserved by designStructure preservation for JSON, CSV and code is a stated capability of the models.
Your next move · Explorer

Test constraints, not fluency

1Take a real document with a glossary and a strict output schema, and run the failing cases first.
2Score constraint compliance separately from translation quality, because the two fail differently.
3Check per-language behaviour for the pairs you actually ship, not the 150-language average.

Switch your reading mode at the top to see a different next move.

Tap to open

Things to keep an eye on

Pop quiz · unlock the Constraints as instructions badge

Did it stick?

0/3
What distinguishes this release from a general-purpose translation model?+20 XP
Under what licence was the family released?+20 XP
What is the biggest gap in the evidence for this release?+20 XP
Your call · +5 XP

You are choosing a translation backend for a product.

Research dossierThe full research, labeled and sourcedCOMPANY CLAIM12 sources · 5 min read · 20 sections
Story identity

Story ID: S39
Title: Bilibili open-sources Index-Translate, an Apache 2.0 multilingual translation family covering 150 languages
Organization: Bilibili Index Team
Category: Open-Weight Models / Language AI
Event date: 2026-10-01
Announcement date: 2026-10-01
Article dates: ["2026-10-01", "2026-10-02"]
Evidence status: COMPANY CLAIM
Confidence: Medium
Rank: 39/40
What changed (from discovery): Bilibili's Index team released Index-Translate under Apache 2.0 in three sizes (2B, 9B, 35B-A3B preview), built on Qwen3.5 with ~3B active parameters per token, covering 150 languages with instruction-following translation capabilities.

Primary claim (FACT/COMPANY CLAIM distinction): Bilibili Index Team released Index-Translate under Apache 2.0 license with three model sizes covering 150 languages (COMPANY CLAIM). This is documented in official GitHub repository and HuggingFace model cards.

✓

What happened?

🎓 For Explorer

On 2026-10-01, Bilibili Index Team announced and released Index-Translate, a family of multilingual translation models built on Qwen3.5. The release includes:

  • Index-Translate (text): 2B, 9B, and 35B-A3B preview variants covering 150 languages
  • Index-Echo S2TT/S2ST (speech-to-text/speech-to-speech) in 2B and 9B
  • Index-Homura (syllable-controlled translation) in 2B and 9B
  • Index-NativeLong (document-level translation) in 2B and 9B
  • All released under Apache 2.0 license

The models support instruction-following translation with hard constraints (glossaries, format preservation, code/JSON structure) and soft constraints (tone/style, domain disambiguation). The family is publicly available on GitHub, HuggingFace, and ModelScope.

Δ

What changed?

Before (CONTEXT): Multilingual translation models existed but often with restrictive licenses or limited instruction-following capabilities for constrained translation tasks. High-quality translation models with Apache 2.0 licensing and explicit support for terminology/glossary enforcement were not as widely available in open-weight form at this scale.

Change (EVENT): Bilibili open-sourced Index-Translate family (Apache 2.0) with 150-language coverage, instruction-following translation, speech variants, syllable control, and long-document translation capabilities.

After (IMMEDIATE): Developers can freely use, modify, and distribute these models commercially under Apache 2.0. The family provides multiple variants for different use cases (text, speech, controlled length, long documents).

A machinist's bench split between a hardwood jig whose part fails to fit its cut-out and a thin spring gauge beside a second part that nothing prevents using.
↔

Before → Change → After

🎓 For Explorer
PhaseStateKey details
BeforeLimited open Apache 2.0 multilingual translation with instruction constraintsExisting models had varying licenses; instruction-following for translation constraints (terminology, formatting) less prominent in open releases
ChangeRelease of Index-Translate familyApache 2.0, 150 languages, 2B/9B/35B-A3B sizes, plus Echo/Homura/NativeLong variants
AfterFreely usable open-weight translation familyCommercial use allowed under Apache 2.0; supports hard/soft constraints, speech, syllable control, document translation
⚙

How it works

FACT (based on documentation):

  • Base architecture: Built on Qwen3.5 models (Qwen3.5-2B/9B and a 35B-A3B MoE variant for preview)
  • Active parameters: ~3B active parameters per token for the 35B-A3B MoE (COMPANY CLAIM)
  • Language coverage: 150 languages for the text models (COMPANY CLAIM)
  • Instruction format: Uses instTrans-style canonical prompts with numbered hard/soft constraints (【硬性要求】, 【注意】)
  • Hard constraints: Glossary enforcement, format/structure preservation (JSON/CSV/code), placeholders
  • Soft constraints: Tone/style, domain disambiguation, cross-sentence consistency
  • Variants:
    • Index-Translate: general text translation
    • Index-Homura: syllable-controlled translation
    • Index-NativeLong: long-document translation (up to large contexts)
    • Index-Echo S2TT/S2ST: speech-to-text and speech-to-speech with voice conditioning
!

Why it matters

🎓 For Explorer

The release provides Apache 2.0-licensed multilingual translation with explicit constraint handling. This addresses localization needs for multilingual products, especially where terminology consistency and formatting preservation matter (technical docs, legal text, structured data). Apache 2.0 removes licensing friction for commercial deployment.

✦

What became possible?

🎓 For Explorer
  • Commercial deployment of a 150-language translation family without copyleft/licensing restrictions
  • Terminology-controlled translation for regulated/brand-sensitive content
  • Multimodal translation workflows (text + speech) using open models
  • Syllable-controlled translation for subtitles/dubbing timing
  • Long-document translation with better context handling
  • Local/private deployment of capable translation models
◎

Implications

Technical

COMPANY CLAIMS (to be treated as such):

  • Scale: 2B/9B/35B-A3B models covering 150 languages - broad coverage in open weights
  • Architecture: MoE (35B-A3B) with ~3B active params per token - efficiency trade-off
  • Quantization: GGUF, FP8 (W8A8), FP4 (NVFP4 W4A4) variants released for deployment flexibility
  • Context handling: Large context for NativeLong variant (specific limits not fully detailed in initial release materials reviewed)
  • Instruction-following: Native support for translation constraints integrated into training (as per docs)

INTERPRETATION: This expands the open Apache 2.0 multilingual translation ecosystem, particularly with constraint-aware translation capabilities.

Developer

  • Access: Models available on GitHub, HuggingFace, ModelScope with inference guides
  • Deployment: Supports vLLM, llama.cpp (GGUF), standard transformers
  • Extensibility: Apache 2.0 allows modification/forking
  • Integration: Can be integrated into localization pipelines, subtitle tools, apps
  • Quantized builds: Multiple precision levels for different hardware

Enterprise

  • Cost control: Can self-host instead of API calls; Apache 2.0 commercial-friendly
  • Data privacy: On-premise deployment possible for sensitive content
  • Localization at scale: 150 languages for global products
  • Brand consistency: Glossary enforcement supports terminology control
  • Compliance: Apache 2.0 license clear for enterprise use

Strategic

  • Ecosystem: Strengthens China's open-source LLM ecosystem (Bilibili Index Team)
  • Competition: Competes in open multilingual translation space against other open models
  • Localization democratization: Broad language coverage including lower-resource languages (150 total)
  • Licensing: Apache 2.0 is strategically permissive for adoption
⚠

Risks & limitations

Risks
  • Quality claims unverified: Language coverage and quality metrics are COMPANY CLAIMS; independent multilingual benchmarks not yet widely available from third parties
  • Model size/cost: 35B variant requires significant GPU memory even with quantization
  • Bias/performance variation: Performance likely uneven across 150 languages (common for multilingual models)
  • Maintenance: Long-term support depends on Bilibili Index Team
Limitations
  • Independent verification: No peer-reviewed third-party evaluation identified in discovery window materials
  • Coverage details: Specific low-resource language performance not independently validated
  • Speech variants: Language coverage for speech models differs from 150-language text coverage
  • Context specifics: Exact context length limits for variants need verification from docs
  • Evidence status remains COMPANY CLAIM until independent evaluation appears
?

Open questions

  1. How does quality compare independently to leading multilingual translation models across all 150 languages?
  2. What are the effective context lengths for NativeLong in practice?
  3. How reliable is glossary enforcement across different language pairs?
  4. What is the real-world performance on low-resource languages in the set?
  5. Are there any usage restrictions or additional terms beyond Apache 2.0?
↗

What happens next?

🎓 For Explorer
  • Short-term (days-weeks): Community will evaluate and benchmark; quantized variants already available
  • Medium-term (weeks-months): Independent evaluations likely to appear; potential fine-tunes and integrations
  • Long-term (months): May influence open multilingual translation standards; ecosystem adoption depends on quality/performance
★

Editorial takeaway

🎓 For Explorer

Bilibili's Index-Translate release is a notable contribution to open multilingual translation under Apache 2.0. The combination of 150-language coverage, constraint-aware instruction following, and multiple specialized variants (speech, syllable control, long-document) makes it a versatile open alternative. The key caveat is that quality claims remain company-provided pending independent verification; nonetheless, the permissive license and comprehensive feature set are significant for developers and enterprises seeking self-hostable translation.

Research sources

Primary Sources (10)
Primary
ModelScope collection - **source title/organisation:** IndexTeam Index-Translate Collection (ModelScope) - **URL:** https://modelscope.cn/collections/IndexTeam/Index-Translate** Mirrored collection on ModelScope confirming availability and variants. - **publication/event date:** 2026-10-01 (release) — ** PRIMARY — official mirror
URL unavailable
Primary
Index-Echo S2ST model (HuggingFace) - **source title/organisation:** IndexTeam/Index-Echo-S2ST-9B (HuggingFace) - **URL:** https://huggingface.co/IndexTeam/Index-Echo-S2ST-9B** Speech-to-speech translation variant. - **publication/event date:** 2026-10-01 (release) — ** PRIMARY — official model card
URL unavailable
Primary
Index-Echo S2TT model (HuggingFace) - **source title/organisation:** IndexTeam/Index-Echo-S2TT-9B (HuggingFace) - **URL:** https://huggingface.co/IndexTeam/Index-Echo-S2TT-9B** Speech-to-text translation variant of family. - **publication/event date:** 2026-10-01 (release) — ** PRIMARY — official model card
URL unavailable
Primary
Index-NativeLong (Index-Nailong) model (HuggingFace) - **source title/organisation:** IndexTeam/Index-Nailong-9B (HuggingFace) - **URL:** https://huggingface.co/IndexTeam/Index-Nailong-9B** Model card for document-level translation variant (published as Index-Nailong). - **publication/event date:** 2026-10-01 (release) — ** PRIMARY — official model card
URL unavailable
Primary
Index-Homura model (HuggingFace) - **source title/organisation:** IndexTeam/Index-Homura-9B (HuggingFace) - **URL:** https://huggingface.co/IndexTeam/Index-Homura-9B** Model card for syllable-controlled translation variant, part of family. - **publication/event date:** 2026-10-01 (release) — ** PRIMARY — official model card
URL unavailable
Primary
Index-Translate demo - **source title/organisation:** Index-Translate Demo (Bilibili) - **URL:** https://index-translate.bilibili.com/** Official demo showcasing translation, instruction-following, Homura (syllable control), NativeLong (document), Echo (speech). Demonstrates capabilities described in release. - **publication/event date:** 2026-10-01 (release) — ** PRIMARY — official demonstration
URL unavailable
Primary
Technical report (arXiv) - **source title/organisation:** Index-Translate: A Multilingual Translation Model Family — Text, Speech, Controlled Dubbing, and Long-Document Translation (arXiv) - **URL:** https://arxiv.org/abs/2609.40181** Technical paper describing architecture, training methodology, evaluation, language coverage, variants. Reports metrics and model details including 35B-A3B with ~3B active params per token (COMPANY CLAIM). - **publication/event date:** 2026-09-30 / posted around release timeframe — ** PRIMARY — technical documentation from authors
URL unavailable
Primary
Index-Translate collection (HuggingFace) - **source title/organisation:** IndexTeam Index-Translate Collection (HuggingFace) - **URL:** https://huggingface.co/collections/IndexTeam/index-translate** Official collection of all models in the family (text, Echo, Homura, NativeLong variants) with sizes and quantized builds (GGUF, FP8, FP4). Confirms full family release. - **publication/event date:** 2026-10-01 (release) — ** PRIMARY — official collection
URL unavailable
Primary
Index-Translate-9B model card (HuggingFace) - **source title/organisation:** IndexTeam/Index-Translate-9B (HuggingFace) - **URL:** https://huggingface.co/IndexTeam/Index-Translate-9B** Official model card detailing capabilities, training, evaluation results, inference instructions, prompt format, default settings, and model family. Confirms Apache 2.0 license, Qwen3.5 base, 150 languages for text models, instruction-following with hard/soft constraints. - **publication/event date:** 2026-10-01 (release) — ** PRIMARY — official model documentation
URL unavailable
Primary
GitHub repository - **source title/organisation:** Bilibili Index-Translate (GitHub) - **URL:** https://github.com/bilibili/Index-Translate** Official repository with README, model descriptions, inference code, quantization info, and Apache 2.0 license. Documents 150 languages for text models, variants (Index-Translate, Index-Echo, Index-Homura, Index-NativeLong), sizes (2B/9B/35B-A3B preview), and instruction-following capabilities. - **publication/event date:** 2026-10-01 (release announcement date) — ** PRIMARY — official project repository
URL unavailable
Independent Sources (2)
Independent
GitHub release-related activity / social coverage (reference) - **source title/organisation:** HuggingFace Papers collection for Index-Translate - **URL:** https://huggingface.co/papers/2609.40181** HuggingFace Papers page linking to arXiv paper, corroborating publication. - **publication/event date:** 2026-10-01 — ** INDEPENDENT PLATFORM — corroborates technical paper availability
URL unavailable
Independent
Pandaily coverage - **source title/organisation:** Pandaily - **URL:** https://pandaily.com/bilibili-index-translate-open-source-qwen3-5-150-languages** Independent media coverage reporting Bilibili open-sourced Index-Translate family under Apache 2.0, covering 150 languages across multiple sizes, built on Qwen3.5. Corroborates key details from release. - **publication/event date:** 2026-10-02 — ** INDEPENDENT — third-party reporting corroborating announcement
URL unavailable
Up next · story 40 of 40
China's MIIT signalled that ByteDance and Alibaba may buy Nvidia's RTX Pro 5500
All storiesNext story