Index-translate Model FamilyFrom text to speech,
bring meaning across languages

Index-Translate: Meaning and Instructions, Together

For character dialogue, voice matters. For reader-facing explanations, tone and formatting matter. For internet memes, meaning often goes beyond the literal words. The same sentence can call for more than one good translation.

Index-Translate combines multilingual translation, instruction following, and meme understanding. Alongside the source text and target language, you can specify how the translation should read: preserve forms of address, use a more conversational tone, or return only the translation. The model uses these requirements to produce text suited to its intended use.

Teasing, irony, and cultural references require an understanding of why something is said that way. We include social and cultural context in our evaluations and assess translation instruction following separately, checking both the quality of the translation and whether the requested constraints were met.

Index-Translate is built upon Qwen 3.5 2B/9B/35A3B base and trained through three stages:

Mid-train: At this stage, we adopt a standard WSD scheduler, mixing pre-training replay corpora, multilingual corpora, parallel corpora, and pivot corpora. We explored diverse data ratios, parallel alignment, and parallel organization methods, specifically increasing the proportion of pivot corpora during the annealing phase, enabling the model to learn cross-lingual semantic correspondences and expressions across diverse languages, laying a solid foundation for downstream translation tasks.

Post-train: Building on the multilingual foundation from mid-train, we trained three separate expert models—a general translation model, an instruction-following translation model, and a meme translation model—conducting SFT and RaR RL (Rubric as Reward) or RLVR for each.

Model-merge: Combining linear interpolation, MOPD, and Mix-RL, we consolidated multi-version capabilities into a unified model.

邹衍站在原地说不出的委屈，他本以为两女在门口等她，怎么也会决个胜负，留下一个陪他。
让他也炫耀虚荣一般，难道我就这么命苦吗？
对不起了，昨天电源踢掉了，稿子都没了，尽力补。

Zou Yan stał w miejscu, czując niewypowiedzianą krzywdę. Myślał, że obie kobiety będą na niego czekać przy drzwiach i że w końcu rozstrzygną między sobą, która zostanie z nim.

Chciałby chociaż trochę pochwalić się swoją dumą. Czy naprawdę mam tak pechowe życie?

Przepraszam, wczoraj padło zasilanie i straciłem cały manuskrypt. Postaram się to nadrobić.

请将以下中文小说翻译成波兰语，并且严格遵循所有约束要求。

【源文】
邹衍站在原地说不出的委屈，他本以为两女在门口等她，怎么也会决个胜负，留下一个陪他。
让他也炫耀虚荣一般，难道我就这么命苦吗？
对不起了，昨天电源踢掉了，稿子都没了，尽力补。

【约束要求】
1. 【注意】指代一致性——源文第一句'他本以为两女在门口等她'中的'她'容易误指为两女中的一位，实际应为'他'（指邹衍自己），可能为笔误。翻译成波兰语时需注意代词性别一致，避免歧义。


只输出译文，不要有任何额外说明。

<div class="chat-msg">天呐，这<span class="bad">充电宝</span>太坑了！<br>拿<i>数据线</i>连上半天，电量不增反降，真让人抓狂。<b>电池健康</b>迟早被搞废，还能不能好好玩了？</div>

<div class="chat-msg">세상에, 이 <span class="bad">휴대용 밥통</span> 진짜 최악이잖아!<br><i>생명줄</i>로 반나절이나 연결했는데, 배터리가 차는 건 고사하고 오히려 줄어들다니, 진짜 미치겠네. <b>배터리 컨디션</b>도 조만간 다 망가질 텐데, 이걸로 제대로 놀 수 있긴 한 거야?</div>

请将以下这段中文HTML代码片段翻译成韩文。在翻译过程中，请务必严格保持所有的HTML标签原样不变。此外，请遵循指定的术语映射要求：将“电池健康”译作“배터리 컨디션”，将“充电宝”译作“휴대용 밥통”，并将“数据线”译作“생명줄”。最后，请确保整体译文呈现出日常口语化的表述风格。

<div class="chat-msg">天呐，这<span class="bad">充电宝</span>太坑了！<br>拿<i>数据线</i>连上半天，电量不增反降，真让人抓狂。<b>电池健康</b>迟早被搞废，还能不能好好玩了？</div>

注意仅输出翻译内容，不要其他任何解释

xdm关于e父的喜加一

我想吟下此诗：

静观其变，白嫖就好，入库吃灰，能要就要

敬请期待，e宝开造。

Leute, was das nächste Gratis-Spiel von Epic angeht:

Ich muss da mal dieses Gedicht rezitieren:

Warte einfach ab, hol’s dir umsonst, lass es im Bibliotheksregal stauben – wenn’s geht, nimm’s!

Bleibt dran, Epic legt los!

请将以下中文专栏文章翻译成德语，并且严格遵循所有约束要求。

【源文】
<p>xdm关于e父的喜加一</p><p><br></p><p>我想吟下此诗：</p><p><br></p><p>静观其变，白嫖就好，入库吃灰，能要就要</p><p><br></p><p>敬请期待，e宝开造。</p><p><br></p><p></p><p></p>

【约束要求】
1. 【硬性要求】保留源文中的 HTML 格式标记不变
2. 【注意】消除歧义——文中'白嫖'为网络游戏用语，指不花钱免费获取游戏/资源，而非字面意义上的'嫖娼'；'入库吃灰'指领取后放入库中却不游玩；'e宝'指Epic Games平台（昵称）。这些均需保留原网络黑话风格并意译。
3. 【注意】译文保持口语化网络聊天/游戏社区用语风格，不随意切换语体或语气

只输出译文，不要有任何额外说明。

<p>xdm关于e父的喜加一</p><p><br></p><p>我想吟下此诗：</p><p><br></p><p>静观其变，白嫖就好，入库吃灰，能要就要</p><p><br></p><p>敬请期待，e宝开造。</p><p><br></p><p></p><p></p>

<p>Leute, was das nächste Gratis-Spiel von Epic angeht:</p><p><br></p><p>Ich muss da mal dieses Gedicht rezitieren:</p><p><br></p><p>Warte einfach ab, hol’s dir umsonst, lass es im Bibliotheksregal stauben – wenn’s geht, nimm’s!</p><p><br></p><p>Bleibt dran, Epic legt los!</p><p><br></p><p></p><p></p>

Text and instructions → Translation

ONLINE WORKFLOW

1. Add your text. Choose the source and target languages, then enter the text to translate.
2. Set your requirements. Specify tone, forms of address, or output format. Custom instructions use the 9B version.
3. Read the translation. The translation appears progressively as it is generated, making it easy to read and inspect.

Try Index-Translate · Text translation ↗

Index-Echo: Let Another Language Be Heard

Narration, dialogue, and commentary in video need to reach audiences through sound. Index-Echo takes source speech as input and generates the translation as target-language speech, letting the content be heard in a new language.

Index-Echo is an end-to-end S2S (Speech-to-Speech) translation model supporting speech translation and dubbing in 6 languages. With fast inference for dialogue and video clips, it makes the path from hearing the original to receiving translated speech more efficient.

S2TT: speech in, translated text out. An AuT encoder is connected to Index-Translate-2/9B, and the speech translation task is trained end-to-end.

S2ST: speech in, translated speech out. Building on S2TT, the CosyVoice3 tokenizer is removed, and the LLM's last hidden representations are connected to CosyVoice3's semantic layer. Training proceeds in two stages: first, the LLM and CosyVoice3 are frozen while an approximately 30M-parameter mapper is distilled to align the representations; then the participating model parameters are unfrozen and trained end-to-end with DiffRO. Speaker voice information is extracted from the source audio and used for speech generation; it currently supports translation and dubbing from Chinese/English into 6 target languages.

The current online workflow starts with a Bilibili video and provides translated subtitles, proper-name and terminology settings, and a separate dubbing entry point. You can review the text before proceeding to dubbing.

Video → Subtitles and dubbing

ONLINE WORKFLOW

1. Provide a video. Paste a Bilibili video URL, BV ID, or av ID, and select the target language.
2. Review translated subtitles. Specify proper names and terms as needed, and export bilingual ASS subtitles.
3. Continue to dubbing. Use the separate dubbing action to continue producing the video's voiceover.

Try Index-Echo · Video speech translation ↗

Index-Homura: Translation That Fits the Rhythm of Dubbing

A line spoken in a few seconds can become much longer in another language. Cross-language dubbing therefore needs to do two things at once: preserve the meaning and fit the script into the original time window.

Index-Homura is a translation model with syllable control, designed for dubbing. You can set a target syllable-ratio range so that length becomes part of the translation task. The ratio is the number of translated syllables divided by the number of source syllables; the current online demo specifies length using a target syllable count.

Making length control a training objective. The HOMURA method uses GRPO reinforcement learning to combine a dynamic syllable-ratio reward with a translation-quality reward, jointly optimizing length constraints, meaning preservation, and fluency. The reward interval accounts for differences between target languages and the coarse granularity of counting syllables in short sentences, teaching the model to choose wording that fits the budget.

Dialogue and syllable count → Dubbing script

ONLINE WORKFLOW

1. Enter Chinese dialogue. Choose English, Arabic, Spanish, or Japanese as the target language.
2. Set the target syllable count. Use the slider or numeric input to specify the desired translation length.
3. Review the script. Check its meaning and length before preparing the voiceover.

Try Index-Homura · Syllable-controlled translation ↗

Index-NaiLong: A Whole Novel, in One Pass

Long documents expose a clear weakness in specialized translation models. In this Chinese-to-English evaluation, Hy-MT2-7B falls on GuoFeng from 0.7334 at 4K to 0.0001 at 64K, and on BWB from 0.7167 to 0.0001. By 16K, the two scores have already fallen to 0.0197 and 0.0334. Specialized translation ability does not automatically extend to reliable long-document translation.

Even frontier models can face a serious bottleneck in long-form output. LongWriter (2024) observed that the frontier models evaluated at the time could accept long contexts yet struggled to generate sufficiently long text on request. Translating a whole novel requires sustained source coverage, consistent characters and plot, and a translation that reaches the end. A larger context window alone does not demonstrate these abilities.

Index-NaiLong targets long inputs and outputs and supports generating the complete translation of a novel in one pass. We make full-length translation a core task, examining output completeness, consistency of character names and forms of address, and continuity across passages, with the goal of producing a complete work that can be used further.

In the online demo, you can upload a document, read the translation as it is generated, and download the result. The current demo processes documents in segments and presents Chinese-to-English translations progressively.

Long document → Downloadable translation

ONLINE WORKFLOW

1. Upload a document. Provide Chinese content in TXT or Markdown format.
2. Read the translation in segments. The demo processes the document in chunks and displays English results as they become available.
3. Download the translation. Save the result as Markdown or PDF for further use.

Try Index-NaiLong · Long-document translation ↗

Evaluation

Detailed evaluations for Index-Translate, Index-Echo, Index-Homura, and Index-NaiLong follow. Each model uses test sets and metrics appropriate to its task.

Index-Translate: Text Translation

We organize text translation evaluations into seven categories: instruction following, books, subtitles, social and cultural context, FLORES, WMT, and low-resource languages.

Among the models in the current comparison, the 9B text model ranks first in instruction following and subtitle translation, and in the top two in five categories. The 2B model ranks third in instruction following. These scores apply only to the evaluated text-model checkpoints, not to 35B-A3B, Homura, NaiLong, or Echo. Their mapping to the final Index-Translate release weights remains to be verified.

Values are unweighted means of the raw metrics within each category; higher is better. All seven categories include 13 models.

On individual benchmarks, Index-Translate-9B achieves an instTrans_minor IFscore of 0.8792, while Index-Translate-2B scores 0.7851, ranking first and second among the reported results. On the subtitle benchmark MuST-Cinema , the 9B model scores 0.8979 COMET-22; on MEME , it scores 0.7331, ranking second among the reported results.

Each category is independently normalized to 0–100 to show relative positions. Distances and areas do not represent raw score differences or overall capability.

Rankings reflect numerical order in the current data, without significance testing. For example, the top two subtitle-category means differ by 0.00065; this does not establish a statistically significant lead.

Download 12 raw benchmark scores · Download all seven-category means

Get started

The Index-translate family includes four models. Choose your content type and desired output to open the corresponding demo.