Index-NativeLong-9B v1.1

🌐 Online demo · GitHub · Technical report · ModelScope

New in v1.1: This update merges Index-Translate-9B with the NativeLong long-document model, bringing their text translation capabilities together in one model: continuous long-document translation, general multilingual translation, and translation instructions covering terminology, formatting, style, and context. This update also improves long-document translation quality.

Index-NativeLong-9B, published as IndexTeam/Index-Nailong-9B, is a long-document translation model in the Index-Translate family. It reads the full source text and translates it in a single generation. By keeping the source and translation in the same context, it can connect details across distant passages and maintain consistent names, terminology, and narrative flow.

The model supports a 262,144-token (256K) context window, shared by the prompt, source, and generated translation. The released weights are text-only, BF16, and MTP0, using standard autoregressive decoding without an MTP prediction head.

Supported tasks

Use NativeLong to translate books, articles, and other long texts in context. Version 1.1 also handles everyday text translation and instructions about terminology, formatting, and style. See RELEASE_NOTES.md for the update summary.

The official document client provides fixed templates for Chinese→English, English→Chinese, Chinese→Japanese, and Japanese→Chinese (zh-en, en-zh, zh-ja, ja-zh). These templates ask for complete translation, consistent names and terminology, and preservation of chapter and paragraph order. Version 1.1 also incorporates the multilingual text translation and translation instruction-following capabilities of Index-Translate-9B. You can specify other target languages and requirements such as glossary use, format and structure preservation, style adaptation, and disambiguation using background context. In some long-document translation scenarios, the model can also follow additional instructions, such as using specified terminology.

Inference

Use a CUDA GPU and a vLLM environment with Qwen3.5 text-model support. The BF16 weights occupy approximately 17.9 GB; serving requires additional memory for the runtime and KV cache. Full-context serving may require multiple GPUs.

git clone https://github.com/bilibili/Index-Translate.git
cd Index-Translate
pip install -r inference/llm/requirements.txt

# Serve a downloaded copy of this release in your vLLM environment.
vllm serve /path/to/downloaded-model \
  --served-model-name IndexTeam/Index-Nailong-9B \
  --max-model-len 262144 --host 127.0.0.1 --port 8000

From the repository directory in another terminal, translate a UTF-8 document using the official direction template:

python inference/llm/doc_translate.py novel.zh.txt \
  --direction zh-en --model IndexTeam/Index-Nailong-9B \
  --output novel.en.txt

Replace the filenames with your documents. For the other directions, use en-zh, zh-ja, or ja-zh. The client streams the translation to the output file, uses greedy decoding, and disables thinking. See INFERENCE.md for endpoint, decoding, and output-budget settings. Input and output share the 256K window; reserve enough space for the translation.

NativeLongBench evaluation

NativeLongBench evaluates full-document translation across five length groups: 4K, 8K, 16K, 32K, and 64K Chinese-side tokens. Our main comparison covers 274 passages from GuoFeng, BWB, and General Books. GuoFeng and BWB contain Chinese web fiction translated into English; General Books contains English books paired with model-generated Chinese references. The linked benchmark release provides GuoFeng and BWB metadata, reconstruction information, and evaluation code.

Each model receives the complete passage and translates it in one generation. Document-SEGALE/COMET aligns source and translated passages, then scores translation quality against references. Each corpus score equally weights the five length groups. Scores below are shown on a 0–100 scale; higher is better.

Model	GuoFeng zh→en	BWB zh→en	General Books en→zh
Index-NativeLong-9B v1.1	81.08	77.64	89.79
Qwen3.8 Flash	70.53	68.23	82.06
Hy-MT2-7B	20.82	19.83	23.60
Hy-MT2-30B-A3B	29.04	29.98	33.06

Index-NativeLong-9B v1.1 achieves a three-corpus average of 82.84. Baseline results follow the original long-document evaluation.

Multilingual translation and instruction following

Version 1.1 combines NativeLong's continuous long-document translation with Index-Translate-9B's multilingual text translation and translation instruction following. The comparison below examines how these text translation capabilities are retained after the merge.

We use the main repository evaluation configuration to compare v1.1 with the original Index-Translate-9B model in the same evaluation period. The four benchmarks below cover multilingual translation quality and translation instruction following, with v1.1 performing close to the original model overall.

TI3000 / instTrans covers social posts, articles, fiction, subtitles, and other text types, evaluating both translation quality and instruction following.
FLORES-200 and WMT24++ evaluate multilingual translation between languages and translation from English into multiple target languages, respectively.
IFMTBench evaluates translation under individual or combined requirements for terminology, style, background context, and formatting.

Quality, COMET-22, and xCOMET-XXL measure translation quality; IF measures instruction following. All scores are shown on a 0–100 scale; higher is better.

Benchmark	Metric	Original Index-Translate-9B (same-period evaluation)	Index-NativeLong-9B v1.1
TI3000 / instTrans	Quality / IF	65.60 / 84.56	65.28 / 84.88
FLORES-200	COMET-22	87.81	87.76
WMT24++	COMET-22	85.97	85.85
IFMTBench clean	xCOMET-XXL / IF	79.54 / 84.34	79.20 / 84.41

Evaluation details and sample counts are provided in RELEASE_NOTES.md.

Model family

Index-NativeLong is part of the Index-Translate family, alongside models for general text translation, speech translation, and translation with a target syllable count. Explore the official model collection. NativeLong retains the IndexTeam/Index-Nailong-* repository IDs.

Citation

@techreport{indextranslate2026,
  author={Tianjiao Li and Mengran Yu and Chenyu Shi and Lusheng Zhang and
          Qisi Chen and Yanshan Zhou and Ji Qi and Jingying Liu and
          Yuang Feng and Ziang Cui and Tianxing Yan},
  title={Index-Translate: A Multilingual Translation Model Family --- Text, Speech, Controlled Dubbing, and Long-Document Translation},
  institution={Index LLM Team},
  year={2026},
  month={September}
}