Index-Translate-35B-A3B-preview, released on 30 September 2026 by Bilibili, reads images as well as text and ships a 262,144-position configuration. It suits teams that want glossary-constrained translation they can host themselves, or prototype against on a free public API before committing any hardware.
Can you use it?
- Index-Translate free public API: preview (Free OpenAI-compatible endpoint; a placeholder key is accepted). No rate limit or terms stated on the pages read.
- Index-Translate online demo: preview (Free web demo with text and image input).
- Free plan: yes.
- Open weights: released (Apache-2.0).
Checked 2 Nov against Bilibili's own pages. Compare access across frontier models
What changed
- · New feature Index-Translate-35B-A3B-preview: Index-Translate free public API plans: Free OpenAI-compatible endpoint; the README says any string works as the API key → Free OpenAI-compatible endpoint; a placeholder key is accepted source
- · New feature Image translation added across the family source
- · New feature Free public API and quantized builds source
Index-Translate-35B-A3B-preview is Bilibili's open mixture-of-experts translation model, with 35B total and 3B active parameters. Bilibili reports a 0.8794 FLORES COMET-22 score for this preview, and the weights are free to download and run on your own servers, with a free hosted API as an alternative.
Provider: Bilibili · Family: Index-Translate
Context window: 262,144 tokens
Input modalities: text, image · Output: text
About Index-Translate-35B-A3B-preview
Index-Translate-35B-A3B-preview is a mixture-of-experts translation model from Bilibili's Index team, released on 30 September 2026 as the largest member of the Index-Translate family. It is built on the Qwen3.5 base from Alibaba Cloud, carries 35B total parameters with 3B active per token, and is a preview checkpoint that the team describes as an in-training version. Hugging Face metadata lists the size as 36B, while the model card says 35B-A3B; Hokai uses the card's figure. The 2B and 9B dense siblings ship under the same family name and are tracked here as one family.
The model translates text across 150 languages and follows instructions about terminology, formatting, style and output length, and since 10 October 2026 the whole family also translates text inside images without a separate OCR step. Bilibili reports a FLORES COMET-22 score of 0.8794 and an instruction-following score (instTrans IFscore) of 0.8336 for this preview, and 76.76 on its WMT26 Judge scale. In the same table GPT-5.6 Sol scores 89.10 and DeepSeek V4.1 Flash 83.55 on WMT26 Judge, so the closed and larger models still lead on that axis. These are vendor-reported numbers from Bilibili's technical report, and no independent evaluation exists yet.
Access is open. The weights are Apache-2.0 on Hugging Face, with GGUF, FP8 and NVFP4 builds added on 3 October 2026. Since 4 October 2026 Bilibili also runs a free OpenAI-compatible public API at index-translate.bilibili.com that serves this model. The README says any string works as the API key, and the pages Hokai read state no rate limit or terms for that endpoint, so treat it as a demo service, not a production commitment. For your own servers the card shows a vLLM command with a 32,768-token served context, even though the shipped configuration allows 262,144 positions.
Where it fits: pick it when you need self-hosted translation with glossary and format constraints, or a free endpoint to prototype against. Pick DeepL or a closed API model when top-line quality matters more than control. The family guide at best AI for translation compares the wider field.
Pricing
Weights are free to download, so cost is your own GPU time. Bilibili runs a free OpenAI-compatible API for this model and publishes no price, limit or terms for it. See [Bilibili](/hub/companies/bilibili) for the maker and [DeepL](/hub/tools/deepl) for a paid alternative.
Key Features
- 150-language translation: Follows glossary, style and length instructions across the full language set, per the model card.
- Image translation: Reads text inside images end to end, with no separate OCR step, since the 10 October 2026 update.
- Open weights: Weights on Hugging Face, plus GGUF, FP8 and NVFP4 builds for different hardware.
- Free public API: OpenAI-compatible endpoint at index-translate.bilibili.com, live since 4 October 2026.
Pros
- Three quantized builds make self-hosting realistic.
- Hard and soft constraints keep glossaries, formats and tone under control.
- Free public API removes any setup for a first test.
Cons
- Preview status: Bilibili calls it an in-training version.
- Closed models lead on Bilibili's WMT26 Judge table (89.10 for GPT-5.6 Sol against 76.76).
- No published rate limits, terms or retention policy for the free API.
Frequently Asked Questions
What does Index-Translate-35B-A3B-preview cost to use?
The weights are free under Apache-2.0, so you pay only for your own hardware. A free hosted endpoint from Bilibili serves the model too, and the pages Hokai read state no price, rate limit or terms for it. Closed models such as GPT-5.6 Sol charge per token, so cost is the main reason to try this one.
Is Index-Translate-35B-A3B-preview better than GPT-5.6 Sol at translation?
Not on Bilibili's own tables: GPT-5.6 Sol scores 89.10 on WMT26 Judge against 76.76 for this preview. The Index model stays close on FLORES COMET-22 and is built for glossary and format constraints, so it wins when control and hosting matter more than peak quality. No independent test exists yet.
Can I use Index-Translate-35B-A3B-preview commercially?
The Hugging Face card lists Apache-2.0, which permits commercial use with attribution and the license notice. The model is built on Qwen3.5, so check the base model's terms as well. The free public API is a separate service whose terms Hokai could not find.
Does Index-Translate-35B-A3B-preview keep my text private?
If you download the weights and serve them yourself, text never leaves your servers. If you use the free public API at index-translate.bilibili.com, Hokai found no published retention or training policy, so avoid sending confidential text there.
Who should skip Index-Translate-35B-A3B-preview?
Skip it for production work that needs a service-level guarantee, since the model is a preview and the API has no stated limits. Teams that need the highest translation quality should compare it with a closed API model or DeepL first. It is a strong fit for researchers and engineers building multilingual pipelines on their own GPUs.