Keywords：ERNIE 4.5, Baidu open source, multimodal model, PaddlePaddle framework, Apache 2.0, ERNIE 4.5 performance, MoE model, MFU 47%, DeepSeek-V3 comparison, Qwen3 competition
🔥 Focus
Baidu Releases and Open-Sources ERNIE 4.5 Model Series: Baidu has officially open-sourced its ERNIE 4.5 series, which includes 10 text and multimodal models with parameter sizes ranging from 0.3B to 424B (a MoE model with 47B active parameters). The series is trained on the PaddlePaddle framework, achieving an MFU of 47%, and is licensed under the Apache 2.0 protocol, supporting commercial applications. In multiple benchmarks, its performance surpasses models like DeepSeek-V3 and Qwen3, and its multimodal version outperforms OpenAI o1 in some evaluations. This open-source release not only provides model weights and code but also simultaneously opens up API services, aiming to promote the widespread application of AI technology in both academia and industry. The community has responded enthusiastically, looking forward to it competing with top open-source models like Qwen (Source: huggingface, tokenbender, teortaxesTex, jeremyphoward, QbitAI)