MoE、大规模、多模态、长文本，Llama4系列模型开启开源AI新纪元

Investment Rating - Investment advice: Llama 4 marks a new era in open-source AI, expected to boost the ecosystem [2][5] Core Insights - Llama 4 series models, including Llama 4 Scout and Llama 4 Maverick, were released by Meta, featuring massive parameters and superior performance compared to competitors [2][5] - Llama 4 Scout has 17 billion active parameters and 109 billion total parameters, outperforming models like Gemma 3 and Gemini 2.0 Flash-Lite [2][5] - Llama 4 Maverick, with 17 billion active parameters and 400 billion total parameters, excels in image localization and surpasses GPT-4o in benchmarks [2][5] - Llama 4 Behemoth, currently in training, will have 288 billion active parameters and 2 trillion total parameters, expected to outperform GPT-4.5 and Claude Sonnet 3.7 in STEM benchmarks [2][5] - The models utilize Early Fusion technology for multimodal capabilities, integrating text and visual tokens into a unified framework [2][5] - Llama 4 Scout supports a context window of up to 10 million tokens, setting a record for open-source models [2][5] - The Llama series is anticipated to drive further development in the open-source AI ecosystem, following the success of Llama 2 and the upcoming Llama 3 [2][5] Company Focus - Recommended targets include: Empyrean Technology, Dameng Data, China National Software & Service, Beijing Kingsoft Office Software, Hygon Information Technology, Cambricon Technologies, Dawning Information Industry, Ninestar Corporation, Isoftstone Information Technology, Newland Digital Technology, Guangzhou Sie Consulting, Jiangsu Tongxingbao Intelligent Transportation Technology, and Range Intelligent Computing Technology Group [2][5]