Foxconn's logo is displayed during the Hon Hai Tech Day at the Nangang Exhibition Center in Taipei, Taiwan, Oct. 8, 2024. AP-Yonhap
Foxconn's logo is displayed during the Hon Hai Tech Day at the Nangang Exhibition Center in Taipei, Taiwan, Oct. 8, 2024. AP-Yonhap
Foxconn Technology Group, the world's largest electronics contract manufacturer and major iPhone supplier for Apple, launched its first Chinese large language model (LLM) trained on traditional characters, as the Taiwanese company pushes forward the use of artificial intelligence (AI) in factories.
The new FoxBrain model was trained in a "more efficient and lower-cost" method within just four weeks, and sets a new milestone in the development of Taiwan's AI technology, according to a statement issued on Monday by Foxconn, known formally as Hon Hai Precision Industry.
With a training process powered by 120 Nvidia H100 graphics processing units (GPUs), FoxBrain excels in math and logical reasoning, according to Foxconn.
It was originally designed for internal applications in the company, but Foxconn said it will be open sourced in the future, as part of efforts to collaborate with technology partners to expand its applications and promote AI in manufacturing.
The Nvidia's GPU (Graphic Processing Unit) is shown in this photo taken in Paris, Febr. 23, 2024. AFP-Yonhap
The Nvidia's GPU (Graphic Processing Unit) is shown in this photo taken in Paris, Febr. 23, 2024. AFP-Yonhap
LLMs are the technology underpinning generative AI services like OpenAI's ChatGPT. Open source gives public access to a software's source code, allowing third-party developers to modify or share its design, fix broken links or scale up its capabilities.
Foxconn's latest initiative reflects the company's goal to push its own AI breakthroughs in terms of manufacturing efficiency.
That follows Chinese start-up DeepSeek's launch earlier this year of its high-performance R1 reasoning model, which was open-sourced and developed at a fraction of the cost of AI models from larger companies like OpenAI, Google and Meta Platforms.
"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," said Li Yung-Hui, director of the Artificial Intelligence Research Centre at Hon Hai Research Institute, the research arm of the manufacturer, in the statement.
"Our FoxBrain model adopted a very efficient training strategy, focusing on optimizing the training process rather than blindly accumulating computing power," he said.
The new model was based on the Meta Llama 3.1 architecture with 70 billion parameters. Foxconn claimed that it outperformed Llama-3-Taiwan-70B, another open-source model fine-tuned on traditional Chinese characters and English data using the Llama-3 architecture, in most categories of TMMLU+, a benchmark for traditional Chinese language understanding.
Models developed by Chinese companies like DeepSeek are generally trained for better understanding of simplified Chinese characters, which are used on the mainland.
Last November, the company said it was working with Nvidia to leverage "digital twin" technology in manufacturing and supply chain management.
The initiative uses Nvidia's Omni verse to streamline global factory operations, enhance resilience and ensure consistent quality.
Read the full story at SCMP.