• 📢 News • 🤗 Open Resources • 📄 Papers
The LMT project aims to advance the frontier of Multilingual Machine Translation (MMT) by building Inclusive, Scalable, and High-performance multilingual translation models.
- 2026.4.18: Our RouteLMT paper RouteLMT: Learned Sample Routing for Hybrid LLM Translation Deployment has been accepted to ACL 2026 Industry Track (Oral). RouteLMT learns to route translation samples among LLM-based MT models of different sizes, enabling more efficient hybrid deployment.
- 2026.4.7: Our LMT paper NiuTrans.LMT: Toward Inclusive and Scalable Multilingual Machine Translation with LLMs has been accepted to ACL 2026 Main Conference. This work introduces a family of strong LLM-based MT models covering 60 languages and 234 translation directions.
- 2025.11.11: Our LMT paper NiuTrans.LMT: Toward Inclusive and Scalable Multilingual Machine Translation with LLMs and corresponding Models are released.
We have made the following resources available:
| Resource | Description | Link |
|---|---|---|
| LMT-60-*B | Our high-performance multilingual translation models cover 60 languages and 234 directions. Available in four sizes: 0.6B / 1.7B / 4B / 8B. | LMT-60-0.6B LMT-60-1.7B LMT-60-4B LMT-60-8B |
| LMT-60-*B-Base | Our continued pre-training of Qwen3 on 90B tokens serves as the foundation for large-scale translation adaptation. Available in four sizes: 0.6B / 1.7B / 4B / 8B. Note that these models have not been fine-tuned for translation using our SFT data. For translation tasks, please refer to the LMT-60-*B series. | LMT-60-0.6B-Base LMT-60-1.7B-Base LMT-60-4B-Base LMT-60-8B-Base |
| LMT-60-sft-data | Our SFT dataset including Flores-200 devset, NTREX-128, SMol, WMT14–23, and IWSLT17–24 test sets, totaling 567K samples. | LMT-60-sft-data |
| FLORES-mvf | An Inner Mongolian (mvf) evaluation set annotated by native speakers to extend the FLORES-200 benchmark. | FLORES-mvf |
LLM-based MT models achieve strong translation quality, but using the largest model for every input is costly in deployment. RouteLMT (routing for LLM-based MT) addresses this with hybrid deployment, where translation samples are routed among LLM-based MT models of different sizes under a fixed budget. RouteLMT learns a lightweight in-model router from the small translator's prompt-token representations to estimate the expected gain of using a larger model before generation. By assigning large-model calls to the most beneficial samples, RouteLMT improves the quality–budget trade-off for efficient LLM-based MT deployment.
In this project, we take a step toward overcoming the prevailing English-centric bias in MMT. We introduce LMT, a suite of Chinese-English-centric MMT models trained on 90B mixed monolingual and bilingual tokens, covering 60 languages across 234 translation directions and achieving SOTA performance among models with similar language coverage. Our work makes the following main contributions:
- We identify and analyze a previously overlooked issue, directional degeneration, in large-scale multilingual SFT with multi-way data and propose a simple yet effective Strategic Downsampling method to mitigate it.
- We propose Parallel Multilingual Prompting (PMP), which enhances cross-lingual transfer by incorporating an auxiliary parallel sentence into the instruction.
- We release LMT, a suite of large-scale Chinese–English-centric multilingual translation models in four sizes (0.6B/1.7B/4B/8B), providing strong baselines for future MMT research.
| Resource Tier | Languages |
|---|---|
| High-resource Languages (13) | Arabic(ar), English(en), Spanish(es), German(de), French(fr), Italian(it), Japanese(ja), Dutch(nl), Polish(pl), Portuguese(pt), Russian(ru), Turkish(tr), Chinese(zh) |
| Medium-resource Languages (18) | Bulgarian(bg), Bengali(bn), Czech(cs), Danish(da), Modern Greek(el), Persian(fa), Finnish(fi), Hindi(hi), Hungarian(hu), Indonesian(id), Korean(ko), Norwegian Bokmål(nb), Romanian(ro), Slovak(sk), Swedish(sv), Thai(th), Ukrainian(uk), Vietnamese(vi) |
| Low-resource Languages (29) | Amharic(am), Azerbaijani(az), Tibetan(bo), Modern Hebrew(he), Croatian(hr), Armenian(hy), Icelandic(is), Javanese(jv), Georgian(ka), Kazakh(kk), Central Khmer(km), Kirghiz(ky), Lao(lo), Inner Mongolian(mvf), Marathi(mr), Malay(ms), Burmese(my), Nepali(ne), Pashto(ps), Sinhala(si), Swahili(sw), Tamil(ta), Telugu(te), Tajik(tg), Tagalog(tl), Uighur(ug), Urdu(ur), Uzbek(uz), Yue Chinese(yue) |
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "NiuTrans/LMT-60-8B"
tokenizer = AutoTokenizer.from_pretrained(model_name, padding_side='left')
model = AutoModelForCausalLM.from_pretrained(model_name)
prompt = """Translate the following text from English into Chinese:
English: The concept came from China where plum blossoms were the flower of choice.
Chinese:"""
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
generated_ids = model.generate(**model_inputs, max_new_tokens=512, num_beams=5, do_sample=False)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()
outputs = tokenizer.decode(output_ids, skip_special_tokens=True)
print("response:", outputs)For more details, please refer to src/inference.py.
For questions, please contact: luoyingfeng_neu@outlook.com
If you find our paper useful for your research, please kindly cite our paper:
@misc{luo2026routelmt,
title={RouteLMT: Learned Sample Routing for Hybrid LLM Translation Deployment},
author={Yingfeng Luo, Hongyu Liu, Dingyang Lin, Kaiyan Chang, Chenglong Wang, Bei Li, Quan Du, Tong Xiao, Jingbo Zhu},
year={2026},
eprint={2604.22520},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2604.22520},
}
@misc{luoyf2025lmt,
title={NiuTrans.LMT: Toward Inclusive and Scalable Multilingual Machine Translation with LLMs},
author={Yingfeng Luo, Ziqiang Xu, Yuxuan Ouyang, Murun Yang, Dingyang Lin, Kaiyan Chang, Tong Zheng, Bei Li, Peinan Feng, Quan Du, Tong Xiao, Jingbo Zhu},
year={2025},
eprint={2511.07003},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2511.07003},
}