Index-Translate: A Multilingual Translation Model Family -- Text, Speech, Controlled Dubbing, and Long-Document Translation

Tianjiao Li; Mengran Yu; Chenyu Shi; Lusheng Zhang; Qisi Chen; Yanshan Zhou; Ji Qi; Jingying Liu

cs.CLSep 30, 2026

Index-Translate: A Multilingual Translation Model Family -- Text, Speech, Controlled Dubbing, and Long-Document Translation

Authors: Tianjiao Li, Mengran Yu, Chenyu Shi, Lusheng Zhang, Qisi Chen, Yanshan Zhou, Ji Qi, Jingying Liu, +3 more

Organizations: Index LLM Team

Abstract

We introduce Index-Translate, a multilingual translation model family that combines a shared multilingual foundation with specialized training for general translation, instruction following, speech translation, controlled dubbing, and long-document translation. It includes three model sizes, 2B, 9B, and 35B-A3B, and supports translation in 150 languages, with multilingual instruction following. Evaluations on general translation and complex translation instructions show that Index-Translate outperforms translation models of comparable size and achieves performance comparable to 100B-scale translation models and frontier models. Image-text evaluation shows that translation remains possible without image examples in Index-Translate-specific training, although results vary by model size. Index-Echo provides end-to-end speech-to-text and speech-to-speech translation, outperforming existing end-to-end models and achieving performance comparable to frontier omni models. Index-Homura extends the family to syllable-controlled dubbing. Index-NativeLong introduces native long-document translation with a dedicated task formulation and benchmark. These capabilities support diverse translation tasks, including multilingual content production.

Figures & tables

Model	Backbone	Task
Index-Translate	2B / 9B / 35B-A3B	Multilingual text and image-text translation, with translation instructions
Index-Echo S2TT	2B / 9B	Speech-to-text translation
Index-Echo S2ST	2B / 9B	Speech-to-speech translation with voice conditioning
Index-Homura	2B / 9B	Translation with a specified syllable count
Index-NativeLong	2B / 9B	Native long-document translation

Table 1: Index-Translate model family and evaluated model sizes.

Figure 2: Training overview of the text-model family, including 2B, 9B, and 35B-A3B. General, Instruction, and Meme specialists share a multilingual mid-trained initialization. Parameter interpolation is followed by targeted MOPD, using domain-matched teachers to supervise student-generated rollouts with a reverse-KL objective, to obtain the final models. The displayed interpolation weights apply to the evaluated 2B and 9B models. Judge-based rewards use RIVAL supervision as described in Section 2.2 .

Stage	Sampling ratio	Steps	Tokens (B)	Learning rate
Constant	general : parallel : monolingual $=1{:}1{:}1$	7,000	117.44	$5\times 10^{-5}$
Decay	general : pivot-20 : pivot-148 $=1{:}4{:}2$	3,000	50.33	$5\times 10^{-5}\to 5\times 10^{-6}$
Total		10,000	167.77

Table 2: Base translation mid-training recipe shared by 2B, 9B, and 35B-A3B. Token budgets use 4,096 tokens per sequence and 4,096 sequences per global batch.

Figure 3: Index-Echo speech-to-text and speech-to-speech architecture. S2ST training aligns the mapper and then applies DiffRO; the S2TT backbone remains frozen throughout.

Model	FLORES	WMT24++	WMT26	instTrans		IFMTBench		Vertical	MEME
	COMET-22	COMET-22	Judge	Quality	IFscore	XCOMET-XXL	IFscore	mean
Index-Translate-35B-A3B (preview)	0.8794	0.8586	76.76	0.6901	0.8336	0.7926	0.8991	0.8438	0.7405
Index-Translate-9B	0.8789	0.8601	75.35	0.6771	0.8209	0.7957	0.8760	0.8451	0.7387
Index-Translate-2B	0.8655	0.8489	60.26	0.5391	0.7569	0.7712	0.7584	0.8377	0.6443
Hy-MT2-1.8B	0.8522	0.8401	49.35	0.3181	0.4932	0.7493	0.7161	0.8314	0.3643
Hy-MT2-7B	0.8747	0.8593	60.51	0.5143	0.6079	0.8049	0.8741	0.8335	0.5139

Table 4: Text translation results for Index-Translate and comparison systems. Instruction judging uses GPT-5.6-Sol; quality and IFscore are separate metrics. WMT26 judge scores use a 0–100 scale, off-target rates are percentages, and other scores use 0–1. Vertical is the equal-weight mean of five domain scores. North-Small-Translate denotes the 218B-A25B model, and TranslateGemma denotes translategemma-12b-it . Panel (b) reports low-resource general translation and instruction following. Bold marks the best value in each column: lower is better for off-target rates, and higher is better for all other metrics.

Model	OPUS-Books	Biomedical	MTNT	MuST-Cinema	GuoFeng
Index-Translate-35B-A3B (preview)	0.8013	0.8681	0.8244	0.8962	0.8291
Index-Translate-9B	0.8010	0.8674	0.8246	0.8979	0.8345
Index-Translate-2B	0.7845	0.8635	0.8132	0.8930	0.8345
Hy-MT2-1.8B	0.7773	0.8592	0.8080	0.8838	0.8286
Hy-MT2-7B	0.7954	0.8671	0.8245	0.8929	0.7876
Hy-MT2-30B-A3B	0.8054	0.8686	0.8279	0.8966	0.8310

Table 5: Vertical-domain translation results (higher is better): COMET-22 for the four sentence-level benchmarks and document-COMET for GuoFeng-Webnovel. Biomedical denotes the WMT Biomedical Translation Task. The equal-weight mean of these five columns is the Vertical score in Table 4 . Bold marks the best score in each column.

Probe	Size	Index	Qwen3.5	$\Delta$	Paired 95% CI
MMTIT	2B	0.5204	0.6826	$-0.1622$	$[-0.1705,-0.1538]$
MMTIT	9B	0.7347	0.7667	$-0.0320$	$[-0.0378,-0.0262]$
MMTIT	35B-A3B	0.7845	0.7807	$+0.0038$	$[-0.0005,+0.0079]$
FLORES-rendered	2B	0.5104	0.5221	$-0.0117$	$[-0.0223,-0.0012]$
FLORES-rendered	9B	0.6780	0.6714	$+0.0066$	$[-0.0018,+0.0147]$
FLORES-rendered	35B-A3B	0.7050	0.6926	$+0.0124$	$[+0.0036,+0.0217]$

Table 6: Image-text translation COMET-22. Qwen3.5 columns use post-trained models; the Index 35B-A3B checkpoint is a preview.

\Delta

is Index minus Qwen3.5. Paired bootstrap intervals cluster by image for MMTIT and Vistra and by source language for synthetic FLORES-rendered images. Vistra uses an adapted whole-image protocol; neither supplementary probe reports an official source-benchmark score.

Model	C-Eval	GPQA-Diamond	INCLUDE	MMMLU
Index-Translate-35B-A3B (preview)	77.6	49.1	68.8	71.9
Index-Translate-9B	69.6	36.3	63.1	66.6
Index-Translate-2B	42.6	27.4	39.1	42.7
Hy-MT2-1.8B	46.0	25.7	37.6	40.1
Hy-MT2-7B	57.4	30.5	50.3	52.3
Hy-MT2-30B-A3B	78.0	48.9	67.4	70.3

Table 7: General-capability accuracy (%) on matched evaluation subsets. Index-Translate uses the same model versions as the text-translation results. INCLUDE and MMMLU are averaged over examples; MMMLU samples 300 questions per language. Bold marks the highest score per column.

Model	MT judge $\uparrow$	ASR error (median) $\downarrow$	Start-time MAE (s) $\downarrow$
Qwen3.8-Omni-Flash	0.887	0.112	1.004
Index-Echo-9B	0.857	0.102	0.488
Gemini-3.1-Pro (thinking)	0.849	0.169	1.824
Qwen3.8-LiveTranslate	0.833	0.186	–
Index-Echo-2B	0.825	0.088	0.395
Gemini-2.5-Flash (thinking)	0.817	0.201	2.402

Table 8: S2TT comparison on the in-house video translation test set. The (thinking) label identifies baselines evaluated with thinking enabled. Dashes denote unavailable timestamp outputs. SeamlessM4T-v2 lacks line-level timestamp instructions and receives unchunked approximately 57-second inputs, beyond its approximately 30-second effective range.

			Content error $\downarrow$ (mean / median)		Spk. $\uparrow$
Direction / metric	Size	Text judge $\uparrow$	Pipeline	E2E	Pipeline	E2E
zh $\to$ en / WER	2B	0.830	0.0450 / 0.000	0.0709 / 0.029	0.722	0.743
zh $\to$ en / WER	9B	0.840	0.0477 / 0.000	0.0615 / 0.000	0.718	0.739
zh $\to$ es / WER	2B	0.728	0.0811 / 0.070	0.0704 / 0.051	0.747	0.743
zh $\to$ es / WER	9B	0.793	0.0916 / 0.071	0.0773 / 0.057	0.746	0.746
zh $\to$ ja / Kata CER	2B	0.760	0.0575 / 0.034	0.0493 / 0.027	0.777	0.780

Table 9: Matched S2ST comparison on in-house video dubbing test data at 2B and 9B. Pipeline denotes Index-Echo-S2TT + CosyVoice3. Both systems share the same frozen Index-Echo-S2TT model and text translation. Bold marks the lower mean content error within each row. Spk. is source-speaker cosine similarity.

Model	Quality $\uparrow$	Mean $\|d\|$ $\downarrow$	Within $\pm 1$ $\uparrow$	Within 10% $\uparrow$	Slope $\approx 1$
Index-Homura-9B (RL)	0.7863	0.0693	74.42%	81.92%	0.968
Index-Homura-9B (SFT)	0.8581	0.1752	41.39%	45.75%	0.680
Index-Homura-2B (RL)	0.7615	0.0962	54.39%	63.08%	0.947
Index-Homura-2B (SFT)	0.7965	0.1594	37.89%	43.06%	0.672
GPT-5.6-Sol (low)	0.8715	0.3223	42.53%	47.64%	0.891
DeepSeek-V4-Flash (no thinking)	0.8742	0.3368	16.58%	16.36%	0.153

Table 10: Expanded SandGlass evaluation on 3,600 cases per model.

d

is relative syllable-count deviation. Bold marks the best value among the displayed models (slope closest to 1). GPT-5.6-Sol uses the recorded gateway configuration, which may inject additional system prompts.

Target language	Cases	Quality $\uparrow$	Within 10% $\uparrow$
English	900	0.7950	78.67%
Japanese	900	0.8300	86.67%
Arabic	900	0.7661	79.89%
Spanish	900	0.7539	82.44%

Table 11: Index-Homura-9B by target language on SandGlass. Each language includes 300 sentences at three length targets.

Corpus	Original	Books	Train	Valid.	Test	Mean (K)	P25–P75 (K)
GuoFeng	Chinese	140	130	4	6	171.6	56.7–87.7
BWB	Chinese	337	327	4	6	777.2	277.0–1205.4
General Books	English	306	292	3	11	103.2	72.3–127.5

Table 12: Books used in this study. Lengths describe the available original-language text per training book (K = 1,024 tokens); P25–P75 denotes the middle 50% of the length distribution.

GuoFeng (zh $\to$ en)
Model	4K	8K	16K	32K	64K	Mean
Qwen3.8 Flash	$\mathbf{0.7974}$	$0.7956$	$0.7572$	$0.6826$	$0.4939$	$0.7053$
GLM-5.3-Flash	$0.7931$	$0.7957$	$0.7921$	$0.6018$	$0.5444$	$0.7054$
North-Small-Translate-1.0	$0.7952$	$0.7914$	$0.7309$	$0.5272$	$0.2353$	$0.6160$
Hy-MT2-7B	$0.7334$	$0.2872$	$0.0197$	$0.0006$	$0.0001$	$0.2082$
Hy-MT2-30B-A3B	$0.7535$	$0.4566$	$0.1615$	$0.0142$	$0.0664$	$0.2904$

Table 13: Native long-document translation with document-SEGALE/COMET ( max_size=8 ). Mean equally weights the five Chinese-side length groups. Bold marks the highest displayed score in each column. ∗ Some GLM samples were refused; its General Books scores average the remaining samples.

Figure 4: Native long-document translation across Chinese-side length groups, reproducing the point estimates in Table 13 . Higher document-SEGALE/COMET is better; unmatched blocks receive zero, so the score reflects coverage as well as translation quality. ∗ GLM’s General Books scores average non-refused samples. North uses the experimental context extension described in Section 4.7 .

Constant-stage sampling weights ( $G{:}M{:}P$ ), 7k steps
Setting	Translation	11-task NLU
$1{:}1{:}0$	0.7750	0.7655
$1{:}0{:}1$	0.7897	0.7637
$1{:}1{:}1$	0.7914	0.7725
$1{:}1{:}4$	0.7907	0.7644
Decay-stage sampling weights, 16K context, 25.17B tokens

Table 14: Selected mid-training pilots. The constant panel belongs to the eight-setting sweep completed in Appendix A.2 . The parallel/pivot panel uses Qwen3-8B, a 16K context, and a 20-target-language translation evaluation; both runs reach step 10,000 at the same decay budget. Both settings include

G

and

M

; the first contains no ordinary parallel data

P

. Its translation scores use a different evaluation scope from the appendix decay panel. Bold marks panel maxima.

	Translation			11-task NLU
Model	Base	7k	10k	Base	7k	10k
Qwen3.5-2B	0.7015	0.7707	0.7898	0.5920	0.5303	0.5431
Qwen3.5-9B	0.7726	0.7924	0.8007	0.7774	0.7682	0.7712

Table 15: Training-stage progression in the 20+-language development study, using the metrics defined in Section 5.1 .

Instruction specialist, 9B
Metric	SFT	RL	$\Delta$ (RL $-$ SFT)
instTrans quality	0.6320	0.6706	+0.0386
instTrans IFscore	0.7670	0.8604	+0.0934
IFMTBench XCOMET-XXL	0.7804	0.7941	+0.0137
IFMTBench IFscore	0.8662	0.8977	+0.0315
Low-resource instTrans quality	0.4817	0.5056	+0.0239

Table 17: SFT-to-RL changes for the 9B Instruction and General specialists before model integration. Each RL model is compared with its own SFT baseline. Instruction evaluation uses GPT-5.6-Sol. Score differences are absolute; off-target changes are percentage points (pp), where lower is better.

Figure 5: Quality and syllable-control trade-offs on SandGlass. Points reproduce Index-Homura and selected baselines from Table 10 ; the additional 2B point with syllable-reward weight

w=2

comes from the study in this section. Blue and green denote the 9B and 2B Homura models. Solid arrows connect SFT to RL, and the dashed arrow connects the two 2B reward weights. Both axes favor larger values; arrows show changes between evaluated configurations.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Organization	Illustrative alignment	Parallel degree
Parallel ( $P$ )	$x_{\mathrm{zh}}\longleftrightarrow x_{\mathrm{ja}}$	2 languages
Pivot ( $V$ )	$x_{\mathrm{zh}}\longleftrightarrow x_{\mathrm{en}}\longleftrightarrow x_{\mathrm{ja}}$	3 languages

Appendix

Table 22: Illustrative parallel and pivot groups. Arrows denote cross-language alignment; the pivot example shares one English anchor across the Chinese and Japanese variants.

Phase	Sampling weights	Translation	11-task NLU
Constant, 4K, 7k	$G{:}M{:}P=1{:}1{:}2$	0.7903	0.7638
	$G{:}M{:}P=1{:}2{:}1$	0.7898	0.7711
	$G{:}M{:}P=2{:}1{:}1$	0.7873	0.7641
	$G{:}M{:}P=2{:}2{:}1$	0.7879	0.7725
Decay, 4K, 10k	$G{:}M{:}V=1{:}2{:}2$	0.8023	0.7763
	$G{:}M{:}V=1{:}2{:}4$	0.8033	0.7717

Appendix

Table 23: Additional pilot mixtures. Constant rows complete the sweep in Table 14 . Decay rows form a separate 4K study using the eight-target-language, six-component translation mean defined in Appendix A.2 . NLU is the 11-task macro-average defined in Section 5.1 . Compare translation scores within each study.

(a) Decay sequence length, followed by matched short document SFT
Position encoding	4K decay	128K decay	Gain
RoPE	0.6673	0.7201	+0.0528
NoPE	0.6739	0.7234	+0.0495

Appendix

Table 24: GuoFeng 2B ablations using document-SEGALE/COMET ( max_size=8 ) on a 0–1 scale. Panels (b) and (c) start from 128K-decay bases. Each overall score averages the evaluation passages in its own study.

Code(s)	Language / variety	Code(s)	Language / variety
en , eng_Latn	English	zh , zho_Hans	Chinese
de	German	fr	French
es	Spanish	ja	Japanese
ko	Korean	pt	Portuguese
ru	Russian	ar	Arabic
it	Italian	nl	Dutch

Appendix

Table 32: Core languages: 22 entries, including English and Chinese. Alternate identifiers are grouped together.

Code(s)	Language / variety	Code(s)	Language / variety
ukr_Cyrl	Ukrainian	fas_Arab , pes_Arab	Persian
ces_Latn	Czech	ell_Grek	Greek
dan_Latn	Danish	hun_Latn	Hungarian
fin_Latn	Finnish	nob_Latn	Norwegian Bokmål
slk_Latn	Slovak	bul_Cyrl	Bulgarian
bos_Latn	Bosnian	cat_Latn	Catalan

Appendix

Table 33: Additional languages and varieties: 128 entries. Alternate identifiers are grouped together.

Explore similar work

Sep 12, 2026

North Small Translate: Advanced Cost-Effective Translation (Cohere CAT+)

We present North Small Translate, an open-weight, LLM-based machine translation (MT) model with instruction-following capabilities built on the same foundation as Cohere's Command A Plus, a mixture-of-experts architecture with 25 billion active parameters out of 218 billion total parameters. North Small Translate is trained using difficulty sampling to obtain challenging documents and a five-step training protocol combining supervised fine-tuning, direct preference optimization, and online reinforcement learning. We prioritized throughput through a non-reasoning base model and supplemented with optional agentic capabilities to unlock translation quality gains. North Small Translate is trained to perform MT-related tasks, including post-editing and quality estimation, as well as related tasks such as general instruction following. The model achieves top MT performance across 50 languages in the class of models under 1T parameters, with no need to run expensive reasoning at inference time.

Tom Kocmi, Alexandre Bérard, Phil Blunsom +20

Cohere

Machine Translation Language Model Post-Training

Ask PDF

May 28, 2026

Loong: A Human-Like Long Document Translation Agent with Observe-and-Act Adaptive Context Selection

Document-level translation remains one of the most challenging tasks for large language models, which are constrained by limited context windows that impede global cohesion, while simultaneously suffering from redundant contextual information that degrades translation quality. To address this, we propose a human-like long document translation agent called Loong, which leverages a 3E memory module (Essence-Exemplar-Entity) to store summaries, sentence pairs, and entity records as historical context. Instead of passively attending to all history, Loong performs deep reasoning to adaptively identify the optimal context for translation guidance. Loong optimizes its context policy through reinforcement learning, utilizing preference data derived from its own sampled observe-and-act reasoning trajectories. Empirical evaluations demonstrate that Loong achieves substantial translation quality improvements in English

\Leftrightarrow

Chinese, German, and French directions, with average gains of up to 13.0 points across the three evaluation metrics. Furthermore, Loong exhibits strong generalization across domains and robustness against contextual noise, while maintaining remarkable stability in ultra-long document translation. Our code is released at https://github.com/YutongWang1216/LoongDocMT.

Yutong Wang, Xuebo Liu, Derek F. Wong +6

Institute of Computing and Intelligence, Harbin Institute of Technology, Shenzhen · 2NLP2CT Lab, Department of Computer and Information Science, University of Macau · 3Huawei Translation Services Center

Adaptive Retrieval Document-Level MT Machine Translation Memory-Augmented Language Model Agents

Ask PDF

Jun 19, 2026

CAT-Translate: Building Compact Open-Source Models for Japanese-English Translation

Nowadays, large multilingual translation models demonstrate impressive translation capabilities in the machine translation benchmarks. This raises a practical question to the developers: is it worth developing translation models specialized for a particular language pair if you only need to support that language pair? To give an anecdotal answer to this question, we develop a family of small language models (0.8B, 1.4B, 3.3B, and 7B parameters) specialized for Japanese-English bidirectional translation. We employ a two-stage supervised fine-tuning approach followed by Multi-Objective GRPO (Ichihara et al. 2025) to train models on synthetically generated parallel corpora. We evaluate our models on WMT and real-world translation benchmarks across business, legal, medical, financial, and patent domains. While multilingual models achieve strong performance on WMT benchmarks, our compact models outperform them on real-world benchmarks, suggesting the practical utility of developing specialized translation models even in the era of large multilingual models.

Yuu Jinnai

CyberAgent / Tokyo, Japan

LLM Fine-Tuning Small Language Models Machine Translation

Ask PDF