Foxconn’s Hon Hai Unveils FoxBrain, Cutting-Edge Chinese Language Model

On Mar 10, 2025

The Hon Hai Research Institute, a key division of Hon Hai Technology Group (Foxconn), unveiled a groundbreaking traditional Chinese large language model (LLM) named FoxBrain on Monday.

Initially developed for internal group applications, FoxBrain is poised to play a pivotal role in powering Foxconn’s three major platforms: Smart Manufacturing, Smart EV, and Smart City initiatives.

FoxBrain was designed to enhance the group’s internal systems with capabilities in data analysis, decision support, document collaboration, mathematics, reasoning, problem-solving, and code generation.

The institute revealed plans to open-source the model in the future, marking a milestone in Taiwan’s AI advancements.

The development of FoxBrain represents a significant achievement in Taiwan’s AI sector, as it was completed using an efficient and cost-effective training method in just four weeks.

This rapid progress was enabled by 120 NVIDIA H100 GPUs, scaled using NVIDIA’s Quantum-2 InfiniBand network. The training process also benefited from the Taipei-1 Supercomputer and technical support from NVIDIA.

FoxBrain’s architecture is based on Meta’s Llama 3.1 model, featuring 70 billion parameters and a 128,000-token context window.

A unique adaptive reasoning reflection technique was used during training, enabling advanced autonomous reasoning capabilities.

“In recent months, the deepening of reasoning capabilities and the efficient use of GPUs have gradually become the mainstream development in the field of AI.

“Our FoxBrain model adopted a very efficient training strategy, focusing on optimising the training process rather than blindly accumulating computing power,” explained Yung-Hui Li, director of the Artificial Intelligence Research Centre at Hon Hai Research Institute.

He emphasised that resource optimisation and innovative training methods allowed them to develop a local AI model with robust reasoning capabilities.

While the institute acknowledged a slight performance gap between FoxBrain and DeepSeek’s distillation model, they highlighted that FoxBrain outperforms Llama-3-Taiwan-70B in mathematics and logical reasoning.

Detailed results of the model’s performance will be shared at NVIDIA GTC 2025 in a session titled “From Open Source to Frontier AI: Build, Customise, and Extend Foundational Models” on March 20.

FoxBrain’s launch signifies a new chapter in Taiwan’s AI industry and showcases how streamlined training methodologies can deliver cutting-edge results with reduced costs.

The model’s open-source promise further reinforces its potential to influence the global AI landscape.

Analytics India