On April 1st, Baidu PaddlePaddle, China's pioneering independently developed and open-source deep learning platform, launched its state-of-the-art Framework 3.0. Harnessing five groundbreaking technological advancements, including "dynamic-static unified automatic parallelism," this latest iteration significantly speeds up the training and inference of large models. PaddlePaddle Framework 3.0 seamlessly integrates with various mainstream large models, including ERNIE Bot 4.5 and ERNIE Bot X1, and boasts a doubled throughput for the DeepSeek-R1 full-featured version when deployed on a single machine. These technological leaps empower PaddlePaddle to deliver low-latency, high-throughput, and cost-effective inference services.
