Index-NativeLong-9B
Index-NativeLong-9B, published as IndexTeam/Index-Nailong-9B, is the 9B long-document translation specialist in the Index-Translate family. It receives the full source document and produces its translation in a single generation, keeping distant source passages and the translation history together. This supports content coverage, consistent names and terminology, and continuity across chapters or scenes.
The 9B checkpoint has the highest five-group means among the models in the report: 0.7891 / 0.7683 / 0.8848 on GuoFeng / BWB / General Books. The shipped serving configuration supports a 229,376-token context window, shared by the prompt, source, and generated translation. Use the published Nailong model ID in commands.
Training and supported tasks
The family builds on Qwen3.5 with multilingual mid-training and translation post-training. This 9B model continues from the integrated translation model before full-parameter document SFT. Document supervision consists of aligned book passages and complete target translations, with bidirectional training over approximately 4K–64K Chinese-side tokens. Training sequence length reaches 128K in the long-document adaptation; this is separate from the shipped serving context limit.
The technical report studies long-sequence decay and RoPE/NoPE positional-encoding choices in separate 2B ablations. After document SFT, their overall scores are similar, with strengths varying by length. These experiments describe training alternatives rather than a runtime switch in the released checkpoint.
The released document client supplies fixed prompt templates for Chinese→English, English→Chinese, Chinese→Japanese, and Japanese→Chinese (zh-en, en-zh, zh-ja, ja-zh). Its instructions require full coverage, consistent entity names and terminology, and preservation of chapter, paragraph, and content order. These four directions are the packaged interface; the general text model's 150-language coverage does not establish NativeLong performance in 150 languages.
Inference
Use a CUDA GPU and a vLLM build with Qwen3.5 support; the official guide records testing with vLLM 0.29. The bf16 model-memory guide is approximately 8 GB for 2B and 24 GB for 9B, plus KV cache for the chosen context length. Full-context serving can require substantially more memory or tensor parallelism.
git clone https://github.com/bilibili/Index-Translate.git
cd Index-Translate
pip install -U vllm
pip install -r inference/llm/requirements.txt
bash inference/llm/serve_vllm.sh nailong-9b --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
python inference/llm/doc_translate.py paper.en.txt \
--direction en-zh --model IndexTeam/Index-Nailong-9B \
--output paper.zh.txt
Replace the input filenames with your documents. For Japanese, use --direction zh-ja or --direction ja-zh. The client streams the translation into the output file and loads the prompt from inference/llm/prompts/nailong_<direction>.txt.
Released client defaults
| Setting | Default |
|---|---|
| Temperature | temperature=0 (greedy) |
| Top-p / top-k / min-p | top_p=1, top_k=-1, min_p=0 |
| Seed | seed=42 |
| Presence / repetition penalty | presence_penalty=0, repetition_penalty=1 |
| Frequency penalty | Not explicitly set by this client; use the inference server's default |
| Stop token IDs | stop_token_ids=[248044, 248046] |
| Stop strings | Not explicitly set by this client; use the inference server's default |
| Ignore EOS | ignore_eos=False |
| Thinking | Disabled through chat_template_kwargs={"enable_thinking": False} |
| Output budget | max_tokens omitted by default; generation uses the remaining context window |
| Streaming | stream=True; writes each received content delta to the output |
| Chat messages | One user message rendered from the fixed direction template and complete source |
| Serving context limit for this checkpoint | 229376 tokens |
Complete client arguments
These are the actual defaults in doc_translate.py, shared by both checkpoint sizes. The example above selects this card's checkpoint explicitly. Sampling, stop-token, thinking, and streaming settings are fixed in the client; its command-line interface exposes the following controls:
| Argument | Default and behavior |
|---|---|
input | Required UTF-8 source filename; - reads standard input; normalizes CRLF line endings, strips outer whitespace, rejects empty input |
--output, -o | Standard output if omitted; otherwise writes the streamed translation to the specified UTF-8 file |
--direction, -d | Unset; if supplied, selects one of zh-en, en-zh, zh-ja, ja-zh and takes priority over source/target flags |
--source, -s | zh; used to derive the direction when --direction is omitted |
--target, -t | en; used with --source, so the default derived direction is zh-en |
--model, -m | INDEX_MODEL environment variable, otherwise IndexTeam/Index-Nailong-9B; pass IndexTeam/Index-Nailong-9B to select this checkpoint |
--base-url | OPENAI_BASE_URL environment variable, otherwise http://127.0.0.1:8000/v1 |
--api-key | OPENAI_API_KEY environment variable, otherwise EMPTY |
--max-tokens | 0; omits API max_tokens; a positive value sets an explicit output-token cap |
--help, -h | Prints the command-line help |
The serving preset sets --served-model-name IndexTeam/Index-Nailong-9B and --max-model-len 229376; the example binds it to 127.0.0.1:8000. Extra preset arguments pass through to vllm serve, including deployment options such as --tensor-parallel-size. The client bypasses environment proxies for localhost endpoints.
The context window includes the complete rendered prompt, source text, chat-template tokens, and generated translation. A source that fits on its own may leave too little room for a complete translation. Estimate both source and expected output sizes with the model tokenizer, and select an input size that leaves adequate output capacity. --max-tokens adds an explicit output cap when desired; it does not extend the total window.
The 2B and 9B shipped limits are 262,144 and 229,376 tokens, respectively. The report's native benchmark uses a 256K-token context setting; the public 9B serving preset follows its shipped configuration rather than overriding it to 262,144.
See the inference guide, document client, and fixed templates for full endpoint and prompt details.
NativeLongBench evaluation
NativeLongBench evaluates continuous passages from held-out books in three corpora. GuoFeng and BWB contain Chinese web fiction with English translations. General Books contains English originals from literature, history, social science, science and technology, paired with model-generated Chinese translations. Japanese originals with model-generated Chinese translations provide additional training supervision.
The report divides books by work and removes identified cross-corpus copies of test works from training:
| Corpus | Books | Train / validation / test works | Evaluated direction |
|---|---|---|---|
| GuoFeng | 140 | 130 / 4 / 6 | zh→en |
| BWB | 337 | 327 / 4 / 6 | zh→en |
| General Books | 306 | 292 / 3 / 11 | en→zh |
The main comparison evaluates native translation: every model receives the complete test passage and is asked to translate it in one generation. The GitHub evaluation summary reports 274 document windows across five length groups: 4K, 8K, 16K, 32K, and 64K Chinese-side tokens. For GuoFeng/BWB the length is measured on the Chinese source; for General Books it is measured on the Chinese reference translation. K means 1,024 tokens. These are document-length groups, distinct from the total context window; the General Books 64K group corresponds to approximately 68K–82K English input tokens.
SEGALE aligns contiguous source and output blocks before COMET scores them against references. Unmatched blocks receive zero. Scores are averaged within each document, then within each length group; the corpus mean equally weights the five groups. The table uses document-SEGALE/COMET (max_size=8), on a 0–1 scale where higher is better.
The comparison includes both released Index-NativeLong sizes, the 7B and 30B-A3B translation baselines available in the report, and frontier/API systems under the same native-document protocol.
| Model | GuoFeng zh→en ↑ | BWB zh→en ↑ | General Books en→zh ↑ |
|---|---|---|---|
| Index-NativeLong-9B (this checkpoint) | 0.7891 | 0.7683 | 0.8848 |
| Index-NativeLong-2B | 0.6701 | 0.7300 | 0.8651 |
| Qwen3.8 Flash | 0.7053 | 0.6823 | 0.8206 |
| GLM-5.3-Flash* | 0.7054 | 0.6691 | 0.8795 |
| North-Small-Translate-1.0 | 0.6160 | 0.6282 | 0.8350 |
| Hy-MT2-7B | 0.2082 | 0.1983 | 0.2360 |
| Hy-MT2-30B-A3B | 0.2904 | 0.2998 | 0.3306 |
*Some GLM requests were refused; its General Books scores average the remaining samples. North's long-input evaluation uses an experimental extension beyond its supported 16K input/output range. Gemini 3.8 Flash is excluded from this main comparison because its output limit is below the 128K output budget reserved for the longest group. See the report for these comparison conditions.
This checkpoint's results by length are:
| Corpus / direction | 4K ↑ | 8K ↑ | 16K ↑ | 32K ↑ | 64K ↑ | Mean ↑ |
|---|---|---|---|---|---|---|
| GuoFeng zh→en | 0.7836 | 0.8074 | 0.8080 | 0.8084 | 0.7382 | 0.7891 |
| BWB zh→en | 0.7733 | 0.7695 | 0.7790 | 0.7756 | 0.7440 | 0.7683 |
| General Books en→zh | 0.9016 | 0.8771 | 0.8958 | 0.8980 | 0.8514 | 0.8848 |
Source: the report's main native long-document comparison; GitHub evaluation summary.
Limits and practical use
Native generation helps preserve document-level context, but does not guarantee complete coverage, accurate facts, or consistent terminology on every input. The report observes repeated passages in some 2B outputs. Review long translations for omissions, repetition, untranslated spans, and EOS or context-budget truncation before use.
The fixed evaluation sets have also been used during model development. GuoFeng test works are drawn from its public training corpus because the official test set is unavailable, while being held out from this study's training. General Books uses model-generated references. Interpret the scores with these protocol limits; they do not measure every aspect of a publishable book translation.
The report's case studies compare native 9B with glossary-assisted chunked workflows on conceptual distinctions and recurring names. They illustrate document-level benefits on selected passages, without establishing universal superiority for every workflow. The report's Japanese–Chinese table presents external baselines rather than NativeLong checkpoint scores, so no Japanese quality result is claimed here.
The team plans to release shareable portions of NativeLongBench. Benchmark release remains listed as planned in the GitHub repository; this card links to the current protocol and results.
Model family
| Model | Released checkpoints | Task |
|---|---|---|
| Index-Translate | 2B · 9B · 35B-A3B (preview) | Text translation and translation instructions across 150 languages |
| 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 target syllable count |
| Index-NativeLong | 2B · 9B | Full-document translation; released templates support zh↔en and zh↔ja |
The 150-language coverage refers to the Index-Translate text models. Speech and long-document packages have their own language interfaces. 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}
}
License and feedback
Apache-2.0. Please use GitHub Issues for questions and feedback.