Index-Homura-2B

Index-Homura-2B is the 2B syllable-controlled translation specialist in the Index-Translate family. It takes source text together with an explicit target syllable count, then adjusts the translation's wording toward that count while retaining meaning and natural expression. It also natively supports terminology / glossary constraints (--glossary) simultaneously with syllable count control, strictly preserving required terms while hitting the exact syllable target. This is useful when preparing translated lines for dubbing or timed subtitles.

On SandGlass, the released RL checkpoint achieves 63.08% within 10% of the target count, 54.39% within one syllable, and a separate translation-quality score of 0.7615. Length adherence and translation quality measure different aspects of the task; the count is an approximate control objective.

Training and task

The family builds on Qwen3.5 with multilingual mid-training and translation post-training. Index-Homura adapts the 2B text foundation to an explicitly specified target count. Following the HOMURA reinforcement-learning approach described in the technical report, GRPO jointly optimizes translation quality and a length reward measuring deviation from the requested syllable count.

Use this specialist for explicit count control. The general Index-Translate models also follow subtitle-related instructions, but their source-relative subtitle constraint is a separate task. SandGlass evaluates English, Japanese, Arabic, and Spanish targets; these results do not establish equal control precision across every language in the general text model's 150-language inventory.

Inference

Use a CUDA GPU and a vLLM build with Qwen3.5 support; the official guide records testing with vLLM 0.29. The bf16 memory guide is approximately 8 GB for 2B and 24 GB for 9B, plus the KV cache. The released client uses an OpenAI-compatible chat-completions endpoint.

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 homura-2b --host 127.0.0.1 --port 8000

From the repository directory in another terminal:

# Standard syllable count control
python inference/llm/syllable_translate.py \
  "说到底聊天群的规则一句话就能总结" \
  --syllables 14 --target en --model IndexTeam/Index-Homura-2B

# Syllable count control with terminology glossary
python inference/llm/syllable_translate.py \
  "我们一起去看电影吧" \
  --syllables 7 --target en --glossary "电影:cinema" \
  --model IndexTeam/Index-Homura-2B

The official client supports both standard syllable control and syllable control with terminology glossary constraints:

# Standard syllable count control
请将以下文本翻译为英语，译文严格控制在 14 个音节。直接输出翻译结果，不要进行任何解释。

说到底聊天群的规则一句话就能总结

# Syllable control with terminology constraints
请将以下文本翻译为英语，译文严格控制在 7 个音节。
请严格遵循以下术语对应关系进行翻译：
- 电影 -> cinema
直接输出翻译结果，不要进行任何解释。

我们一起去看电影吧

Released client defaults

Complete client arguments

These are the actual defaults in syllable_translate.py, shared by both checkpoint sizes. The example above selects this card's checkpoint explicitly.

The serving preset uses --max-model-len 32768 and --served-model-name IndexTeam/Index-Homura-2B; the example binds the server 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.

len(text) is the Python character count used to choose the output-token budget. The syllable target must be supplied separately with --syllables; it is distinct from max_tokens. Override the output limit with --max-tokens, or sampling with --temperature, if needed. The same defaults apply to both Homura sizes. The temperature-0 setting for Index-Translate text models does not change Homura's released temperature of 0.3.

See the inference guide and syllable client for endpoint options and full usage.

SandGlass evaluation

SandGlass contains 300 subtitle sentences, with 60 each from animation, film and television, travel, gaming, and knowledge. Each sentence is translated into four target languages at three requested lengths, yielding 3,600 cases per model. The central count comes from subtitle duration and a language-specific speaking rate; short and long budgets scale it by 0.75 and 1.25.

The report evaluates local models with greedy decoding and a 512-token output limit. Those benchmark settings differ from the released client's default temperature of 0.3 and adaptive output budget above. API baselines use their recorded settings; GPT-5.6-Sol uses a gateway configuration that may inject additional system prompts.

Translation quality is a reference-free Gemini-2.5-Flash judgment on a 0/0.5/1 scale, averaged across cases. Relative deviation is the absolute syllable-count error divided by the target count; lower is better. The two adherence columns report the proportions within one syllable or 10% of the target. Regression slope compares output counts with target counts across the three-budget groups; a value close to 1 indicates a stronger response to changes in the requested length.

The comparison includes both Index-Homura sizes and their SFT-stage ablations, same-size Qwen3.5 2B/9B baselines, nearby-size Hy-MT2 1.8B/7B and Hunyuan-MT-7B models, and larger/API systems. Every row uses the same SandGlass task and metrics.

Source: the report's expanded SandGlass evaluation and target-language breakdown; full evaluation tables.

Limits and practical use

Syllable control is approximate. Verify the count in the generated line, especially before recording or synthesizing speech. Exact count does not guarantee exact duration, since pronunciation, pauses, and delivery affect timing.

The report shows a quality–control trade-off: the external GPT and DeepSeek baselines score higher on translation quality while Homura achieves stronger count adherence. In a separate 2B reward-weight experiment, increasing the syllable reward raises the within-10% rate from 63.08% to 74.19% while reducing quality from 0.7615 to 0.6900. That ablation is separate from the released 2B checkpoint's row above.

SandGlass release is listed as planned in the GitHub repository; the report provides the protocol and results. Individual demo examples that hit their targets illustrate possible outputs rather than guarantees for new requests.

Model family

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.