#Taiwan #Taipei #AI Technology #FoxBrain #Hon Hai

Hon Hai Research Institute Launches FoxBrain

On March 11, 2025, the Hon Hai Research Institute, backed by the tech giant Hon Hai Technology Group (widely known as Foxconn), announced a significant milestone in AI technology with the launch of FoxBrain, the first large language model (LLM) developed specifically for Traditional Chinese. This initiative not only marks a new chapter in Taiwan's AI development but also showcases an innovative training approach that is both efficient and cost-effective, achieved in just four weeks.

FoxBrain is designed to enhance internal operations within the Hon Hai Group, facilitating functions such as data analysis, decision support, document collaboration, and problem-solving, including code generation. Aiming to cater to Taiwanese users, FoxBrain has been fine-tuned to align with local language patterns, demonstrating exceptional performance in mathematical reasoning tests and logical assessments.

Dr. Yung-Hui Li, director of the Hon Hai AI Research Center, noted the importance of optimizing training processes rather than merely increasing computational power. This methodology reflects a strategic adjustment that focuses on efficient training systems that leverage GPU resources effectively. The training utilized 120 NVIDIA H100 GPUs connected by a high-speed NVIDIA Quantum-2 InfiniBand network, leading to a total computational resource investment of 2,688 GPU days.

In comparison with recent AI models available in the market, FoxBrain not only delivers top-tier reasoning capabilities but also showcases significant advancements particularly in mathematics. FoxBrain utilizes an architecture derived from Meta's Llama 3.1 with 70 billion parameters, outpacing its contemporaries in performance benchmarks, especially in the TMMLU+ dataset, where it excels in logical reasoning.

To achieve this, the research team employed proprietary technologies to enhance data quality across 24 subject categories, resulting in the generation of 98 billion tokens of valuable pre-training data specifically for Traditional Chinese. The model also features an impressive contextual window length of 128,000 tokens, illustrating its capacity to handle complex tasks involving extensive data.

FoxBrain's training regimen emphasizes adaptive reasoning reflection techniques, allowing the model to develop autonomous reasoning skills effectively. In rigorous testing, FoxBrain has exhibited remarkable improvements in mathematical score metrics, distinguished from the performance of similar existing models, thereby establishing a competitive edge in the realm of traditional Chinese large language frameworks.

Despite being initially designed for internal applications, Hon Hai intends to collaborate with tech partners in the future to broaden the model’s deployment across various sectors. The development team plans to share FoxBrain's capabilities in an open-source format, aiming to promote AI adoption in manufacturing processes, supply chain management, and intelligent decision-making.

Hon Hai's development journey for FoxBrain encapsulated comprehensive stages including data collection, cleansing, increase, supervised fine-tuning, and LRAIF (Reinforcement Learning from AI Feedback). Each phase was meticulously executed, allowing Taiwan's technological talents to develop capabilities that compete on a global scale.

The substantial technological backing from NVIDIA, through the Taipei-1 supercomputer, was instrumental in facilitating FoxBrain's model training success utilizing NVIDIA NeMo. This advancement is set to propel Foxconn's three core platforms—smart manufacturing, electric vehicle technology, and smart city solutions—into the future.

The world will get its first look at FoxBrain’s capabilities during the pivotal NVIDIA GTC 2025 conference session titled 'From Open Source to Frontier AI Build, Customize, and Extend Foundation Models' scheduled for March 20.

About Hon Hai Research Institute

The Hon Hai Research Institute comprises five dedicated research centers, each staffed with an average of 40 high-tech R&D professionals. Their collective focus is on advancing new technologies, bolstering Foxconn's innovation pipeline, and enhancing the corporation’s competitive strategy, referred to as '3+3'.

For further information about Foxconn and its initiatives, please visit here.