谷歌DeepMind近日推出了一款名为Gemma 2 2B的小型AI模型，它在性能和效率方面取得了显著的平衡。这款模型是从更大的Gemma 2 27B模型蒸馏而来，尽管其参数数量只有2.6B，但在LMSYS竞赛中得分超过了GPT-3.5和Mixtral 8x7B，在MMLU和MBPP基准测试中分别取得了56.1和36.6的优异成绩。Gemma 2 2B的发布标志着小型AI模型在业界受到的重视，它可以在多种设备上运行，包括iPhone 15 Pro，并通过NVIDIA TensorRT-LLM优化加速。此外，Gemma 2 2B还集成了多种开发工具，支持各种平台部署，并可以用于研究和商业用途。同时，谷歌还推出了基于Gemma 2构建的安全内容分类器ShieldGemma，以及可解释性工具Gemma Scope，进一步展示了谷歌在AI安全性和透明性方面的努力。这些新工具和模型的发布，无疑将推动AI技术的发展和应用。
英语如下：
News Title: “Google’s New Model Gemma 2 2B Outperforms GPT-3.5 in Benchmarks, Demonstrating the Power of Compact AI”
Keywords: Gemma 2B, Surpasses GPT-3.5, Compact AI
News Content: Google’s DeepMind has recently introduced a small AI model called Gemma 2 2B, which has achieved a notable balance in performance and efficiency. This model is distilled from the larger Gemma 2 27B model, despite having only 2.6 billion parameters, it scored above GPT-3.5 and Mixtral 8x7B in the LMSYS competition, achieving outstanding results of 56.1 and 36.6 in the MMLU and MBPP benchmarks, respectively. The release of Gemma 2 2B signifies the growing importance of compact AI models in the industry, capable of running on various devices such as the iPhone 15 Pro, and accelerated through NVIDIA TensorRT-LLM optimizations. Additionally, Gemma 2 2B integrates multiple development tools supporting deployment across various platforms, and can be utilized for both research and commercial purposes. Furthermore, Google has also launched a security content classifier named ShieldGemma based on the Gemma 2 framework, as well as an explainability tool called Gemma Scope, further demonstrating Google’s efforts in AI security and transparency. The release of these new tools and models is sure to drive the advancement and application of AI technology.
【来源】https://www.ithome.com/0/785/581.htm
Views: 7