GlotMAX series models start from Qwen3 instruct models with layer-slective tuning using small amount of parallel data alone.
Meanwhile, comprehensive testing on 16 reasoning tasks, such as bbeh, Livecodebench, Olymmath and so on, shows that it surpasses existing translation-enhanced models and performs on par with Qwen3 instruct models.
🔥 Excellent Translation Performance
Qwen3-XPlus significantly boost translation performance in both high- and low-resource languages.
image
🔥 Excellent Reasoning Performance
image
Trained Data Covered Languages
en (English)
ar (Arabic)
bn (Bengali)
cs (Czech)
de (German)
es (Spanish)
fr (French)
hu (Hungarian)
ja (Japanese)
ko (Korean)
ru (Russian)
sr (Serbian)
sw (Swahili)
te (Telugu)
th (Thai)
vi (Vietnamese)
zh (Chinese)
Model Index
We implement multiple versions of the Qwen3-XPlus model, the model links are as follows:
If our model helps your work, please cite this paper:
@misc{gaoLLaMAX2YourTranslationEnhanced2025,
title = {{{LLaMAX2}}: {{Your Translation-Enhanced Model}} Also {{Performs Well}} in {{Reasoning}}},
shorttitle = {{{LLaMAX2}}},
author = {Gao, Changjiang and Huang, Zixian and Gong, Jingyang and Huang, Shujian and Li, Lei and Yuan, Fei},
year = {2025},
month = oct,
number = {arXiv:2510.09189},
eprint = {2510.09189},
primaryclass = {cs},
publisher = {arXiv},
doi = {10.48550/arXiv.2510.09189},
archiveprefix = {arXiv}
}
