传媒：Gemma 2发布，降本提效延展AI生态

Investment Rating - The report maintains an investment rating of "Leading the Market" for the industry, indicating an expected investment return exceeding the CSI 300 Index by more than 10% over the next six months [1]. Core Insights - The report highlights the release of Gemma 2 by Google, which features two model sizes: 9 billion parameters and 27 billion parameters. The upcoming 2.6 billion parameter model is expected to be more suitable for AI applications on smartphones. The continuous iteration of large language models is anticipated to accelerate AI empowerment in both B-end and C-end applications [1]. - Gemma 2 demonstrates significant performance improvements over its predecessor, with a notable reduction in deployment requirements, now only needing an NVIDIA H100 Tensor Core GPU or TPU host. The architecture has been comprehensively improved, utilizing local sliding window attention and global attention mechanisms [1]. - The training data for Gemma 2 includes 130 trillion tokens for the 27B model and 80 trillion tokens for the 9B model, sourced from various documents, code, papers, and scientific articles. The model employs knowledge distillation methods to replicate superior performance at a smaller scale [1]. - The report emphasizes that Gemma 2 has shown outstanding performance in several benchmark tests, significantly outperforming previous versions and other standard models in tasks involving complex cognitive challenges [1]. - The investment suggestion indicates that the continuous iteration of high-quality models is expected to expand the AI ecosystem, providing diverse empowerment across various industries. The report recommends focusing on companies such as Tencent, NetEase, Meitu, and others [1]. Summary by Sections Investment Rating - The industry is rated as "Leading the Market" with a maintained rating [1]. Key Events - The report discusses the global release of Gemma 2 by Google on June 27, which includes models with 9 billion and 27 billion parameters [1]. Model Performance - Gemma 2 shows a significant performance increase compared to its predecessor, with deployment requirements greatly reduced [1]. - The model's architecture has been improved with local and global attention mechanisms [1]. Training Data - The training data for the models consists of 130 trillion tokens for the 27B model and 80 trillion tokens for the 9B model [1]. Benchmark Performance - Gemma 2 has excelled in various benchmark tests, showcasing its capabilities in complex cognitive tasks [1]. Investment Recommendations - The report suggests focusing on companies like Tencent, NetEase, and others for potential investment opportunities [1].