Two Models Place in the AIEC Top 10 and Three in the Top 20. In-House Training Results Are Deployed Across Five Key Sovereign AI Sectors.
Taiwan AI Cloud Corporation (Taiwan AI Cloud; stock code: 7735) announced that its Formosa Foundation Model (FFM) series delivered strong results in the September 2026 official language model evaluation conducted by the Administration for Digital Industries, Ministry of Digital Affairs, under the AIEC program. FFM-Gemma4-31B and FFM-Muse-Glimmer-30B ranked ninth and tenth, respectively, in the overall standings. Among models with fewer than 35 billion parameters whose model sizes are explicitly disclosed on the public leaderboard, they ranked second and third. FFM-Gemma4-26B also placed in the overall top 20. All three models, each with approximately 30 billion parameters, scored highly on the Taiwan Values metric and improved their overall performance following in-house training, demonstrating sovereign AI capabilities that combine a strong understanding of Taiwan with model performance.
Peter Wu, CEO of Taiwan AI Cloud, said, “The core of Sovereign AI is enabling governments and enterprises to maintain control over their models, data, deployment, and governance. Self-developed models are a key link connecting compute, data, and applications within a Sovereign AI architecture. As AI becomes part of core workflows in finance, healthcare, judicial services, government, and enterprises, Taiwan needs model capabilities that can be continuously trained, adapted to industry regulations and governance requirements, and deployed in secure, controlled environments. In this AIEC official evaluation, Taiwan-trained models achieved strong results while competing with numerous large international models. This shows that, for Taiwan-specific language contexts and application needs, model size and capital investment are not the only determining factors. The results also reflect Taiwan AI Cloud’s years of commitment to in-house model training and deployment in real-world environments.”
AIECValidates Local Understanding, Strengthening Compliance and Governance for Sovereign AI
Taiwan AI Cloud's three FFM models scored 0.98, 0.96, and 0.94, respectively, on the Taiwan Values metric, representing clear improvements over their corresponding base models before training. Some models improved by more than 70%. The Ministry of Digital Affairs evaluates models' understanding of Taiwan-specific contexts across dimensions including local language, social culture, and values. It has also noted that the location of the servers hosting a large language model determines which jurisdiction's laws may apply. Fields such as finance, legal services, education, and healthcare increasingly require models to reflect Taiwan-specific institutions, regulations, professional knowledge, and practical scenarios.
When selecting AI models, governments and enterprises need to consider more than general capabilities. They must also assess where data is located, whether models can be adapted to local regulations and corporate policies, and whether deployment, access controls, and audit mechanisms meet governance requirements. As of May 2026, AIEC had tested a cumulative total of 131 domestic and international models. ADI has noted that AIEC evaluations provide businesses and users with a more concrete basis for selecting AI products, while public results can help build trust in commercial partnerships, government procurement, and international markets. The Ministry has also emphasized that Sovereign AI involves the legal jurisdiction governing the location of model servers, and is promoting vertical models for finance, legal services, education, and healthcare that reflect Taiwan's institutions, regulations, professional knowledge, and real-world practices.
More Than Local Understanding: In-House Training Improves Agent Capabilities and Deployment Flexibility
Compared with large language models containing hundreds of billions of parameters offered by overseas platforms, the two FFM models that entered the AIEC top 10 each have approximately 30 billion parameters and require substantially fewer GPU, memory, power, and server resources. This can help reduce the barriers to on-premises or in-country deployment for governments and enterprises, supporting a better balance among model capability, data security, inference efficiency, and long-term operating costs. Taiwan AI Cloud’s training also strengthens reasoning, tool use, and Agentic AI capabilities. All three FFM models improved over their original base models in the AIEC overall evaluation, with the largest gain approaching 20%. According to Taiwan AI Cloud’s PinchBench testing, the trained Gemma4 model improved its Agent task performance by more than 10%, with gains in instruction following, Skill invocation, multi-Skill collaboration, external tool use, and document and data processing.
Deployed Across Five Key Sovereign AI Sectors, Backed by Years of Model Development and Full-Stack Expertise
Taiwan AI Cloud has deployed its FFM series in Sovereign AI projects across finance, healthcare, judicial services, government, and manufacturing. Since launching the Formosa Foundation Model in 2023, its Model Foundry training services have expanded to support open model families including Gemma, Muse, Llama, and Mistral, covering large language models, multimodal models, and embedding models. By integrating AI compute, data governance, model training and management platforms, AFS Suite, and Agent applications, Taiwan AI Cloud provides Full-Stack AI Foundry capabilities spanning compute, data, models, and applications. This helps governments and enterprises build Sovereign AI architectures they can control, govern, and continuously evolve. Learn more:https://www.twcloud.ai/en_us/news/model-foundry/
Notes on Data: AIEC (www.aiec.org.tw) rankings are based on the September 2026 language model evaluation, with data updated through October 2, 2026. Agentic AI test data was compiled onSeptember 30, 2026, using PinchBench / OpenClaw Agent Benchmark to assess Agent tasks, tool use, and multi-step reasoning.
▲ Taiwan AI Cloud's Sovereign AI models rank No. 2 and No. 3 among models with fewer than 35 billion parameters on the Ministry of Digital Affairs' AIEC public leaderboard. Two models rank in the AIEC top 10, with three in the top 20.