Hon Hai Research Institute has unveiled FoxBrain, Taiwan’s first Traditional Chinese large language model (LLM), marking a significant milestone in the country’s artificial intelligence advancements. This innovation strengthens Taiwan’s position in AI-driven manufacturing, supply chain management, and decision-making.
FoxBrain was developed using a streamlined training process, reducing costs and completing development in just four weeks. Unlike conventional models that require extensive time and computational resources, FoxBrain’s efficiency allows for broader AI adoption within Foxconn’s ecosystem.
Built to support data analysis, decision-making, document collaboration, coding, and mathematics, the model is currently integrated into Foxconn’s internal systems. However, the company plans to make FoxBrain open-source, allowing broader access for developers and researchers.
FoxBrain leverages 120 Nvidia H100 GPUs, interconnected through Nvidia’s Quantum-2 InfiniBand, a high-speed networking system crucial for AI training.
Foxconn collaborated with Nvidia’s Taipei-1 Supercomputer, which provided the computational power and expertise needed to refine the model. FoxBrain was developed using Nvidia’s NeMo framework, a toolkit designed to build and customize advanced AI models.
Performance Comparison: FoxBrain vs. Llama-3 Models
FoxBrain is based on Meta’s Llama 3.1 architecture and integrates 70 billion parameters, enabling it to process complex tasks efficiently. Foxconn states that its AI model outperforms Llama-3-Taiwan-70B, a similarly structured Traditional Chinese AI model, in key areas.
Testing shows that FoxBrain excels in mathematical reasoning and logical problem-solving, surpassing Taiwan Llama, which was previously recognized as the top-performing Traditional Chinese language model.
Foxconn applied data augmentation techniques, generating 98 billion tokens across 24 diverse topics to enhance model training. In AI, tokens represent the units of text the system processes, determining the model’s comprehension and accuracy.
With a 128,000-token context window, FoxBrain maintains a broader conversation memory, improving its ability to handle long-form discussions and complex document analysis compared to models with smaller context limits.
Future Developments: Open-Source AI with Expanding Applications
Foxconn acknowledges that while FoxBrain still lags behind DeepSeek’s distillation model in certain areas, it continues to improve and move toward global AI standards.
The AI’s development process included several key stages:
- Data collection and preprocessing
- Data augmentation for enhanced learning
- Continual pre-training and supervised fine-tuning
- Reinforcement Learning from AI Feedback (RLAIF)
Additionally, Foxconn implemented Adaptive Reasoning Reflection, a method designed to refine the model’s analytical and reasoning capabilities.
With plans to make FoxBrain open-source, Foxconn aims to contribute to the growing field of Traditional Chinese AI, providing businesses and researchers with a high-performance language model tailored for complex applications.
