Building, Extending, & Customizing Foundation Models: FoxBrain's Journey to Industrial Frontier

Foxconn's AI Research Center presented FoxBrain, their in-house large language model, at the OCP Global Summit 2025 in San Jose. The talk covers Foxconn's strategic pivot from hardware manufacturing to AI platform builder, the four-component FoxBrain development loop (data engine, model, cluster, eval), and the multi-stage training pipeline including continue pre-training, synthetic data generation, and RL-based post-training. FoxBrain 70B achieves top benchmark results in Traditional Chinese and bilingual evaluations, and is deployed in internal RAG-based office workloads and coding agents.

Key Takeaways

Foxconn is transforming from a Manufacturing Service Company into a Platform Service Company by building three AI-powered smart platforms (Smart Manufacturing, Smart EV, Smart City), all underpinned by the FoxBrain foundation model.
FoxBrain's training infrastructure is a 100 MW Foxconn data center using NVIDIA H100, GB200 NVL72, and GB300 NVL72 GPUs, leveraging tensor, pipeline, and data parallelism with configuration tuned per-cluster.
The FoxBrain data engine applies LLM-as-judge quality control and data enhancement (COSTAR-style rewriting) to curate training data, with a growing shift toward synthetic data — Gartner projects synthetic data will dominate AI training by 2030.
Foxconn's multi-agent post-training data pipeline uses GPT-4+SymPy, CodeLlama, Claude, Gemini, and DeepSeek agents in specialized roles to generate and verify math and coding training samples with quality exceeding individual models.
BeyondWeb synthetic pretraining allows a 3B model to outperform an 8B baseline at the same compute budget (63.7% vs 60.8% avg accuracy), with up to 7.7x training speedup for 8B models over RedPajama.
FoxBrain 70B achieved #1 on Taiwan MMLU+ (66-subject Traditional Chinese benchmark) and outperformed Meta-Llama-3.1-70B-Instruct on Taiwan-MTBench in math (85%), STEM (56.7%), and reasoning (78.3%).
FoxBrain is deployed internally in a hybrid Document-RAG + Graph-RAG office assistant and a coding agent, demonstrating practical industrial applications of the foundation model.