According to DeepLearning.AI, researchers have demonstrated that large language models (LLMs) can be trained using 4-bit FP4 floating-point precision without any loss in accuracy compared to traditional methods. By applying FP4 to matrix multiplications, which account for 95% of training computations, the models achieved performance on par with those trained using the widely adopted BF16 format. This advancement in AI model training reduces computational resource requirements and energy consumption, offering significant cost savings and scalability for enterprise AI deployments. The successful use of FP4 precision presents a new business opportunity for hardware and cloud providers aiming to optimize AI workloads and support more sustainable, large-scale training processes (source: DeepLearning.AI, May 31, 2025).