Cai Fu Zai Xian·2025-06-30 07:40Core Insights - Baidu's Wenxin 4.5 series models have been officially open-sourced on GitCode, providing accessible solutions for enterprises and developers [1][3] - The models include a total of 10 variants, featuring a mixed expert (MoE) architecture with parameter scales of 47B and 3B, and a dense parameter model of 0.3B, with the largest model totaling 424B parameters [3][4] - The MoE architecture allows for cross-modal knowledge integration while retaining dedicated parameter spaces for individual modalities, enhancing multi-modal understanding capabilities [3][4] Model Performance and Features - The Wenxin 4.5 models utilize the PaddlePaddle deep learning framework, achieving a model FLOPs utilization (MFU) of 47% during pre-training [4] - These models have reached state-of-the-art (SOTA) performance across various text and multi-modal benchmark tests, excelling in instruction adherence, world knowledge retention, visual understanding, and multi-modal reasoning tasks [4] - Model weights are open-sourced under the Apache 2.0 license, facilitating academic research and industrial applications [4] GitCode Platform Overview - GitCode, launched on September 22, 2023, has rapidly grown to over 6.2 million registered users and 1.2 million monthly active users, becoming a significant open-source community [5] - The platform integrates advanced code hosting services, supporting version control, branch management, and collaborative development, enhancing the developer experience [5] - The deep integration of Wenxin models with GitCode is expected to drive innovation and sustainable development in the AI industry and the broader open-source ecosystem in China [5] Community Engagement - Ongoing community activities, such as the GitCode × CSDN Wenxin model practical evaluation and discussion series, aim to facilitate developers' understanding and utilization of Wenxin models [6]