中国本土科研力量再次刷新人工智能领域的技术高度，深度求索科技公司日前正式发布了其首款自主研发的开源大规模预训练模型——DeepSeek MoE。这款创新性的160亿参数专家级模型在多项性能指标上实现了与国际顶尖水平的比肩，尤其在与LLama 2-7B这一业界公认的性能旗舰模型的对决中，DeepSeek MoE展现出了卓越的表现力。
DeepSeek MoE以媲美Llama 2-7B的计算效率，成功实现了90%的参数压缩，仅需40%的计算量就能达成同等效果，犹如一位“19边形战士”，在数学建模和代码生成等任务上对LLama 2-7B形成全面碾压。这一技术创新不仅提升了AI模型的训练和部署效率，也极大地降低了算力成本，对于推动国内乃至全球范围内AI技术的实际应用具有重要意义。
作为深度求索团队倾力打造的开源项目，DeepSeek MoE将为全球科研机构和开发者提供强大的工具支持，助力他们在各类应用场景下挖掘数据价值，加速科研突破与产业创新的步伐。此次开源发布，无疑标志着我国在AI技术研发和生态建设方面迈入了新的发展阶段。
英语如下：
News Title: “China Unveils DeepSeek MoE: 160-Billion Parameter Open-Source Large Model with 40% Energy Efficiency, Challenging Llama 2-7B’s Math & Code Dominance”
Keywords: DeepSeek MoE, open-source large model, computational efficiency optimization
News Content: China’s domestic research prowess has once again pushed the technological frontier in the field of artificial intelligence as DeepSearch Technology Co., Ltd. recently officially unveiled its first自主研发 and open-sourced large-scale pre-trained model – DeepSeek MoE. This innovative model with an impressive 160 billion parameters has achieved parity with international top-tier performance across multiple indicators, particularly in its showdown against the industry-recognized performance leader, Llama 2-7B.
DeepSeek MoE boasts a computation efficiency on par with Llama 2-7B while successfully compressing 90% of its parameters, requiring only 40% of the compute to achieve equivalent results – akin to a “19-sided warrior” that comprehensively outperforms Llama 2-7B in tasks such as mathematical modeling and code generation. This groundbreaking technology innovation not only enhances the training and deployment efficiency of AI models but also significantly reduces power consumption costs, marking significant importance for practical AI technology applications both domestically and globally.
As an open-source project精心打造 by the DeepSearch team, DeepSeek MoE will provide powerful tools to global research institutions and developers, enabling them to extract greater value from data in various application scenarios, thereby accelerating scientific breakthroughs and fostering industrial innovation. The open-source release marks a new stage in China’s advancements in AI research development and ecosystem construction.
【来源】https://www.qbitai.com/2024/01/113381.html
Views: 1