Foxconn Unveils FoxBrain—Chinese AI Model Poised for Open Source Release

Hon Hai Research Institute, a research centre by Hon Hai Technology Group (Foxconn), launched a traditional Chinese large language model (LLM) code-named FoxBrain on Monday. 

FoxBrain was initially designed for internal group applications, but it is set to become an important engine for Foxconn’s three major platforms: Smart Manufacturing, Smart EV, and Smart City.

It was initially developed for the Group’s internal systems, supporting applications in data analysis, decision support, document collaboration, mathematics, reasoning, problem-solving, and code generation. The institute also mentioned that the model will be open-sourced and shared publicly in the future.

The institute noted that the development was a milestone in Taiwan’s AI technology, where an efficient and lower-cost model training method was completed in just four weeks.

FoxBrain’s quick training process was made possible using 120 NVIDIA H100 GPUs, scaled with NVIDIA Quantum-2 InfiniBand network. NVIDIA  helped during the training process through the Taipei-1 Supercomputer and technical consultation.

The model is based on Meta’s Llama 3.1 architecture with 70B parameters and features 128k tokens for its context window length. It is said that a unique adaptive reasoning reflection technique was used to train the model in autonomous reasoning.

“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,” said Yung-Hui Li, director of the Artificial Intelligence Research Centre at the Hon Hai Research Institute. “Through carefully designed training methods and resource optimisation, we have successfully built a local AI model with powerful reasoning capabilities.”

The institute mentioned that its performance is close to world-leading standards, but there is a slight gap with DeepSeek’s distillation model.

The institute stated that it outperforms Llama-3-Taiwan-70B on the same scale, with mathematics and logical reasoning as the focus. However, the result of FoxBrain is scheduled to be shared at NVIDIA GTC 2025 in a session talk titled ‘From Open Source to Frontier AI: Build, Customise, and Extend Foundational Models’ on March 20.