计算机行业跟踪报告：谷歌再更新Gemini大模型，立足MoE架构性能更加卓越

Investment Rating - The industry investment rating is "Outperform the Market" (maintained) [1] Core Insights - Google has launched the Gemini 1.5 Pro version, which has performance levels comparable to the largest model, Gemini 1.0 Ultra. The Gemini 1.5 Pro significantly outperforms the Gemini 1.0 Pro in most benchmark tests, with better results in 27 out of 31 tests. It also surpasses Gemini 1.0 Ultra in over half of the benchmark tests, particularly excelling in text and some visual benchmarks [1][9] - The Gemini 1.5 model is built on a sparse mixture-of-expert (MoE) architecture, allowing for more efficient training and service compared to traditional transformer models. This architecture divides the model into smaller "expert" neural networks, enhancing efficiency and reducing computational resource requirements [1][10][12] - Gemini 1.5 Pro features an exceptionally large context window, capable of processing up to 1,000,000 tokens, which equates to handling extensive information such as 1 hour of video, 11 hours of audio, over 30,000 lines of code, or more than 700,000 words. This capability enables complex reasoning over large datasets [1][13][17] Summary by Sections 1. Gemini 1.5 Pro Release - Google has updated its Gemini series with the Gemini 1.5 Pro version, which is a multimodal model capable of processing various types of information including text, images, audio, video, and code [7][9] 1.1 Performance Comparison - The performance of Gemini 1.5 Pro is on par with Gemini 1.0 Ultra, showing significant improvements over Gemini 1.0 Pro in most benchmarks [9] 1.2 Efficiency of MoE Architecture - The MoE architecture allows for more efficient training and service, utilizing a structure that includes sparse MoE layers and a gating network to direct tokens to specific experts [10][12] 1.3 Context Window Capacity - The model supports a context window of 128,000 tokens by default, with options for select users to access up to 1,000,000 tokens, facilitating complex reasoning tasks [13][17] Investment Recommendations - The large context window and efficient MoE architecture present significant investment opportunities, particularly as these models accelerate their application in various fields and continue to drive demand for computational power [19]