浙商证券：近期国产大模型密集发布 规模化应用拉动推理需求

Core Insights - The recent surge in the release of domestic large models indicates the commencement of an AI arms race, with significant advancements in capabilities and applications [1] - The availability of agents is increasing, transitioning large models from chat-based interactions to collaborative tasks, with notable improvements in multi-modal applications [2] - The demand for inference power is expected to rise significantly as large models are applied on a larger scale, particularly in video production and agent functionalities [3] Group 1: Recent Developments in Large Models - Domestic large models have been released intensively, including DeepSeek's new model with a context processing capability of 1M tokens, significantly higher than the previous maximum of 128K [1] - GLM-5 has been launched on the Zhipu website, focusing on programming and agent enhancement, outperforming the latest model Claude Opus 4.6 in global programming tests [1] - ByteDance's Seedance 2.0 has been introduced, which significantly lowers the barriers and costs of video creation, potentially transforming the video production industry [1] Group 2: Advancements in Agent and Multi-Modal Applications - The usability of agents is improving, with models like Claude Opus 4.5 capable of autonomous programming for up to 5 hours [2] - AI coding agents are expected to double their task handling time every 4 months from 2024-2025, a significant acceleration compared to the previous rate of doubling every 7 months from 2019-2024 [2] - Seedance 2.0 supports various combinations of video, audio, and text inputs, producing high-quality video outputs while reducing creation costs [2] Group 3: Inference Demand and Cost Implications - The token consumption for large models is shifting from dialogue and image generation to more intensive applications like agent functionalities and video production, leading to a rapid increase in inference power requirements [3] - The cost of generating a 5-second 720P video is approximately 4 RMB, with Seedance costing around 2.3 RMB, highlighting the significant cost advantages over manual production [3] - The increase in AI penetration in video creation is expected to drive demand for computational power [3] Group 4: Related Companies - Relevant companies include MiniMax-WP (00100), Zhipu (02513), Yunsai Zhili (600602.SH), Youke De-W (688158.SH), Capital Online (300846.SZ), Qingyun Technology-U (688316.SH), Wangsu Technology (300017.SZ), and Nanxing Co. (002757.SZ) [4]