量子位·2025-06-30 04:39Core Viewpoint - Baidu has officially announced the open-source release of the Wenxin large model 4.5 series, providing 10 models with varying parameters and capabilities, including API services for developers [2][4]. Group 1: Model Details - The Wenxin large model 4.5 series includes models ranging from a 47 billion parameter mixture of experts (MoE) model to a lightweight 0.3 billion dense model, addressing various text and multimodal task requirements [2][4]. - The open-source models are fully compliant with the Apache 2.0 license, allowing for academic research and industrial applications [3][14]. - The series features an innovative multimodal heterogeneous model structure that enhances multimodal understanding while maintaining or improving text task performance [5][12]. Group 2: Performance Metrics - The models achieved state-of-the-art (SOTA) performance across multiple text and multimodal benchmarks, particularly excelling in instruction following, world knowledge retention, visual understanding, and multimodal reasoning tasks [9][10]. - In the pre-training phase, the model's FLOPs utilization (MFU) reached 47% [7]. - The Wenxin 4.5 series outperformed competitors like DeepSeek-V3 and Qwen3 in various mainstream benchmark evaluations [10][11]. Group 3: Developer Support and Ecosystem - Baidu provides a comprehensive development suite, ERNIEKit, and an efficient deployment suite, FastDeploy, to support developers in utilizing the Wenxin large model 4.5 series [17]. - The models are trained and deployed using the PaddlePaddle deep learning framework, which is compatible with various chips, reducing the barriers for post-training and deployment [6][15]. - Baidu's extensive AI stack, encompassing computing power, frameworks, models, and applications, positions it as a leader in the AI industry [16].