鸿海扩展自研FoxBrain大模型至VLM/VLA，主打智能制造、电动车与机器人，但仍处内测与小型政务试点阶段

Foxconn is expanding its proprietary FoxBrain large language model (LLM) into vision-language models (VLMs) and vision-language-action models (VLAs), targeting smart manufacturing, electric vehicles (EVs) and robotics. The initiative complements its growing AI server and data centre businesses, with global competition increasingly extending from computing hardware to AI models and applications, alongside rising interest in sovereign AI.

Yung-Hui Li, the founding director of the Artificial Intelligence (AI) Research Center at the Hon Hai Research Institute, said Foxconn launched FoxBrain 1.0 in 2025 after several years of development, followed by version 2.0 in 2026. Beyond the original model's focus on text comprehension, mathematics and logical reasoning, FoxBrain 2.0 strengthens its understanding of Taiwan's languages, history, geography, culture and values to support local applications.
FoxBrain 2.0 ranked 12th among 188 models in a sovereign AI evaluation by the AI Product and System Evaluation Center (AIEC) under Taiwan's Ministry of Digital Affairs. Li said the model remains in internal testing, with a potential open-weight release that would allow developers and businesses to build their own AI services.
Foxconn is also bringing its models into practical use. Its Government Document Brain project, developed with the Yunlin County Government, uses the company's CityGPT platform to draft and review official documents, adjust formatting, and retrieve administrative knowledge, reducing repetitive clerical work. According to Foxconn, the system draws on approximately 20,000 historical documents for model optimisation and validation, using Nvidia GB200 computing resources. The deployment also helps assess the technology's suitability for specialised applications within existing government workflows.
Li said FoxBrain's next phase will introduce visual understanding, enabling VLMs to interpret images and environmental information alongside text. Potential applications include production-line defect detection and intelligent vehicles.
Foxconn then plans to develop VLA models combining vision, language, and action planning with robotic motion-control systems, advancing towards physical AI. Li stressed that models capable of executing physical actions would require safeguards and behavioural constraints to prevent unintended movements.
The shift from LLMs to VLMs and VLAs also aligns with Foxconn's established capabilities in electronics manufacturing, factory automation, EVs and robotics. Integrating proprietary models with production processes, equipment and application platforms could expand AI adoption internally and in customer solutions.
With US-China competition extending into foundation models, Europe and other markets are placing greater emphasis on control over AI models, data and computing resources. Foxconn's capabilities in AI server manufacturing, data centre construction and proprietary model development could allow it to expand beyond hardware supply into customers' AI infrastructure and application integration.