FoxBrain

Overview

FoxBrain is Foxconn's (Hon Hai Research Institute) proprietary foundation model family, with FoxBrain-70B as the flagship model. Built on open-source bases — Meta-Llama, NVIDIA-Nemotron, and DeepSeek — FoxBrain is adapted through a multi-stage training pipeline incorporating domain-specific pre-training, synthetic data augmentation, and reinforcement learning from AI feedback (RLAIF, termed "RL Self-Evol" internally). The guiding principle is "AI Manufactures AI": AI-powered data pipelines generate and curate training data for the next model generation.

FoxBrain is positioned as "Specialized Generalized AI" targeting reasoning, agentic workflow, and robotic and industry solutions. Its development follows a four-block iterative loop: Data Engine (open-source, domain-specific, and synthetic data curation), Model Algorithm (Dense/MoE architectures, RLAIF self-evolution), Cluster (Foxconn's 100 MW data center with NVIDIA H100/GB200/GB300 GPUs), and Evaluation (iterative failure diagnosis and retraining). Continue pre-training injects millions to billions of manufacturing-domain tokens while preserving general capabilities.

As of October 2025, FoxBrain 70B achieved #1 on Taiwan MMLU+ (66-subject Traditional Chinese benchmark covering elementary through professional level), outperforming Taiwan-Llama-70B-Instruct and Meta-Llama-3.1-70B-Instruct. On Taiwan-MTBench (bilingual English-Chinese), FoxBrain-70B-Instruct wins in math (85.0%), STEM (56.7%), and reasoning (78.3%) versus Meta-Llama-3.1-70B-Instruct, using an ensemble of Llama 3.1 405B, Qwen 2.5 72B, and Llama 33 70B as judge models.

FoxBrain is deployed in two internal applications: a hybrid Document-RAG + Graph-RAG office assistant serving Smart EV, Smart Manufacturing, and Smart City knowledge bases, and a coding agent that generates tests, edge cases, and implementation scaffolding.