DeepSeek系列开源模型有望加速AI端侧+应用产业趋势

Investment Rating - The report maintains an "Outperform" rating for the computer industry [2] Core Insights - The DeepSeek series of open-source models, including DeepSeek-V3, R1, and Janus-Pro, exemplifies systematic innovation in algorithm design to achieve efficient utilization of computing power under constraints [4][5] - The training cost for DeepSeek-V3 is approximately $557,000, significantly lower than the estimated $100 million for GPT-4o, indicating a trend towards reduced costs in AI model training [4][5] - DeepSeek-R1 offers a cost-effective API service, charging only 1 yuan per million input tokens, compared to OpenAI's pricing of approximately $15 per million input tokens [6] - The newly released Janus-Pro model demonstrates superior performance in multimodal tasks, outperforming leading models like DALL-E 3 and Stable Diffusion in various benchmarks [7] Summary by Sections Market Performance - The computer industry has shown a market performance fluctuation of -13% to +81% over the specified periods [3] Cost and Efficiency - DeepSeek-V3 reduces hardware resource requirements and training costs, showcasing a significant advancement in distributed inference optimization [4][5] - The model's training efficiency is enhanced through innovative techniques such as auxiliary loss-free load balancing and mixed precision training [5] Performance Metrics - DeepSeek-V3, with 671 billion parameters, competes closely with proprietary models like GPT-4o, achieving high scores in various evaluation metrics [5] - Janus-Pro's multimodal capabilities allow it to excel in both image understanding and generation tasks, achieving a score of 79.2 on the MMBench benchmark [7] Market Implications - The report suggests that the DeepSeek models will accelerate the adoption of AI applications and improve user experiences, leading to a surge in demand for AI capabilities [8] - The anticipated rapid upgrade of edge models and the reduction in inference costs are expected to drive significant growth in the AI sector [8]