Baidu Releases Open Source ERNIE 4.5 AI Models Globally

Chinese tech giant Baidu has open-sourced its ERNIE 4.5 family of AI models, ranging from lightweight 0.3 billion parameter variants to a 424 billion parameter multimodal model, under the Apache 2.0 license. This marks a strategic shift from Baidu's previous closed-source approach, prompted by competition from other Chinese AI firms like DeepSeek, whose open-source models have gained significant international attention. The ERNIE 4.5 models, optimized for language understanding and generation as well as visual modalities, feature a novel heterogeneous Mixture-of-Experts (MoE) architecture designed to enhance multimodal learning and computational efficiency. Benchmark tests show that Baidu's 300B ERNIE 4.5 model outperforms DeepSeek's larger V3 671B model in various tasks, including reasoning, math, and coding, while smaller variants also surpass competing models like Alibaba's Qwen3-30B. Baidu's open-source release includes development tools such as ERNIEKit and FastDeploy, aiming to foster global AI development without proprietary constraints. Industry analysts view this move as a significant contribution to the open-source AI landscape, challenging U.S. and Chinese competitors and signaling a broader industry trend toward openness.