Case Study: Axcxept achieves top Japanese LLM accuracy with Alibaba Cloud Qwen2.5

Axcxept achieves 8.44 average score with Alibaba Cloud's Qwen2.5

Axcxept Inc., a company focused on AI and cloud computing, faced a challenge in the Japanese market where there was no dominant large language model (LLM) specifically tailored for Japanese speakers. They needed a base LLM that worked effectively with the Japanese language and its cultural nuances. They partnered with vendor Alibaba Cloud and utilized its Tongyi Qianwen (Qwen) model as a foundation.

Alibaba Cloud's solution involved using Qwen2.5 as the base for Axcxept's proprietary EZO series model. This fine-tuned implementation resulted in the EZO x Qwen2.5 LLM, which achieved a higher accuracy score than existing closed models on the Japanese MT Bench evaluation. The model also outperformed GPT-4-Turbo in areas like coding, math, and reasoning with an average score of 8.44. Furthermore, the solution is lightweight, offers low latency, and can run in a secure local environment, making it suitable for industries with high security requirements.

Axcxept Inc