Taiwan’s sovereign AI models break into global top 20, ministry evaluation finds

Taiwan’s Ministry of Digital Affairs says locally developed AI models have entered the top 20 in an evaluation designed to test whether artificial intelligence truly understands Taiwan, rather than simply responding in Traditional Chinese.

The ministry announced the results at a press conference titled “Sovereign AI Evaluation and Trusted AI Action.” Through the AI Product and System Evaluation Center, it assessed models in three areas: Taiwanese language, social and cultural understanding, and Taiwanese values. The evaluation used questions from the past five years of Taiwan’s high school entrance examinations in Chinese and social studies, along with material intended to reflect areas of social consensus.

“

Sovereign AI has two important meanings.

Minister Lin Yi-ching said sovereign AI has two important dimensions. Models must follow the laws of the country where their servers are located, and models trained in different countries often carry that country’s viewpoints and ideology. He said many existing models struggle with Traditional Chinese, local Taiwanese knowledge and Taiwanese perspectives, and may apply another country’s political narratives, legal systems or social context.

“

Now we are using questions designed to test people to test these models.

The latest assessment covered 188 models, including nine from Taiwan. Three companies placed models in the top 20. Yating Intelligence ranked fourth, scoring 86.20% in Chinese, 84.00% in social studies and 100% on Taiwanese values. Asia Pacific Intelligence’s ACE-1-24B-2604 ranked 10th, with scores of 83.78%, 71.57% and 100%. FoxBrain v2.0, developed by Hon Hai Research Institute, ranked 12th, scoring 78.70%, 80.51% and 92%.

Lin said the results showed Taiwan could train capable models, including relatively small models competing against systems with larger parameter counts. The ministry plans to expand testing into legal, financial, nursing and education applications, and may add questions from national civil service examinations. Companies can submit models free of charge, while research institutes and universities may later contribute evaluation questions.

“

These good scores are the best marketing tool.