Yating Intelligence’s FedGPT placed fourth overall in the Ministry of Digital Affairs’ latest sovereign AI benchmark, the highest finish for a Taiwan-trained large language model behind three US hyperscaler systems, according to results the government published in early September. Digital Affairs Minister Lin Yi-ching framed the score as proof that 30-billion-parameter models can compete on localized exams without matching trillion-parameter frontier systems on raw scale.
What the test measured
The AI Evaluation Center scored 188 models on language fluency, social context, and values alignment using items drawn from Taiwan’s college entrance-style Chinese and social studies material. FedGPT also posted strong marks on those subject batteries while remaining small enough for on-premise deployment, Yating said in a Taipei briefing.
Foxconn’s FoxBrain and Taiwan AI Cloud’s Llama derivative placed lower but still inside the top fifty, giving policymakers a spread of domestic options as they push agencies to procure locally auditable AI.
Why banks care
Financial supervisors have urged lenders to document model risk when generative tools touch customer service or compliance workflows. Several Taiwanese banks began closed-network pilots with FedGPT-class stacks in late summer, preferring appliances that keep transaction logs inside Taiwan rather than routing prompts to US SaaS defaults.
Yating markets an agentic platform that runs without mandatory cloud egress, a selling point after phishing waves targeted finance employees in September. None of the banks named public success metrics yet; procurement cycles typically lag benchmark headlines by a quarter.
Sovereign AI politics
Lin tied sovereign AI to compute residing under Taiwanese law, repeating that models hosted in jurisdictions without democratic speech protections fail the ministry’s values tests. Opposition lawmakers asked whether rankings favor vendors close to the administration; MODA responded that participation is voluntary and vendors control disclosure.
For enterprise buyers in Taipei, the practical takeaway is narrower: a fourth-place finish is less important than passing the values axis and fitting inside existing security review templates. If October RFPs cite the MODA table, FedGPT’s on-prem story may finally meet budget holders outside innovation labs.