LLaMAX/GlotMAX-101-14B-LST
GlotMAX-101-14B-LST is a 14 billion parameter language model developed by LLaMAX, based on the Qwen3-14B architecture, with a 32K token context length. It is specifically optimized for strong multilingual capabilities, supporting 101 languages, while maintaining robust reasoning performance. The model excels in translation tasks, demonstrating significant improvements over its base model, and performs comparably to Qwen3 instruct models on various reasoning benchmarks.
Loading preview...
GlotMAX-101-14B-LST: Multilingual Translation and Reasoning
GlotMAX-101-14B-LST is a 14 billion parameter language model developed by LLaMAX, building upon the Qwen3-14B instruct architecture. This model is distinguished by its powerful multilingual capabilities and strong reasoning performance, achieved through layer-selective tuning with parallel data.
Key Capabilities
- Exceptional Multilingual Translation: Achieves an average spBLEU score improvement of over 6 points compared to Qwen3-14B on the Flores-101 dataset, and 3.5 points over LLaMAX3-8B-Alpaca.
- Broad Language Support: Supports 101 languages, including Afrikaans, Arabic, Chinese, English, French, German, Hindi, Japanese, Korean, Russian, Spanish, and many more.
- Robust Reasoning: Comprehensive testing across 16 reasoning tasks (e.g., bbeh, Livecodebench, Olymmath) shows performance on par with Qwen3 instruct models, surpassing other translation-enhanced models.
- Efficient Fine-tuning: Utilizes layer-selective tuning with a small amount of parallel data.
Good For
- Applications requiring high-quality multilingual translation across a wide array of languages.
- Tasks demanding strong reasoning abilities in a multilingual context.
- Developers seeking a model that combines translation excellence with robust general intelligence.
For more technical details, refer to the associated paper: LLaMAX2: Your Translation-Enhanced Model Also Performs Well in Reasoning.