近日，全球人工智能研究领域迎来重大突破，专注于通用人工智能（AGI）探索的DeepSeek AI公司宣布开源其最新研发的混合专家（MoE）语言模型——DeepSeek-V2。这款模型以其训练成本低、推理效率高的特性，迅速引起了业界的广泛关注。
DeepSeek-V2拥有236亿个参数，每个token能够激活21亿参数，支持长达128K token的上下文长度，展现了强大的语言处理能力。在性能测试中，DeepSeek-V2在AlignBench基准上的表现超越了GPT-4，且接近于被誉为顶级性能的GPT-4-Turbo，彰显了其在自然语言处理领域的卓越性能。
在多任务基准测试MT-Bench上，DeepSeek-V2与知名的LLaMA3-70B旗鼓相当，同时在与Mixtral 8x22B的比较中展现出优势，尤其在处理数学问题、代码理解和复杂推理任务时，其性能表现令人印象深刻。
这一开源项目不仅为研究者和开发者提供了新的工具，也为AGI的发展开辟了新的道路。DeepSeek AI的这一创新举措，预示着人工智能模型的可访问性和实用性将达到新的高度，有望推动AI技术在各领域的广泛应用。来源：机器之心。
英语如下：
**News Title:** “DeepSeek-V2 Open Source: Low-Cost MoE Model Challenges GPT-4, Stuns the Industry with Its Performance”
**Keywords:** DeepSeek-V2, MoE Model, GPT-4 Challenger
**News Content:**
Title: DeepSeek Unveils Open-Source MoE Model, DeepSeek-V2, Rivaling GPT-4-Turbo, Paving the Way for AGI Milestone
Recently, a major breakthrough in global AI research emerged as DeepSeek AI, a company dedicated to advancing Artificial General Intelligence (AGI), announced the open-source release of its cutting-edge Mixed-Expert (MoE) language model, DeepSeek-V2. The model, known for its low training costs and high inference efficiency, has swiftly captured the attention of the industry.
DeepSeek-V2 boasts 236 billion parameters, with each token activating 21 billion of these, enabling it to handle context lengths of up to 128K tokens, demonstrating its robust language processing capabilities. In performance tests, DeepSeek-V2 outperformed GPT-4 on the AlignBench benchmark and approached the top-performing GPT-4-Turbo, highlighting its exceptional performance in natural language processing.
On the multi-task benchmark MT-Bench, DeepSeek-V2 held its own against the well-regarded LLaMA3-70B and exhibited superiority when compared to Mixtral 8x22B, particularly in solving mathematical problems, code understanding, and complex reasoning tasks, leaving a lasting impression with its performance.
This open-source project not only offers new tools for researchers and developers but also opens up new avenues for AGI development. DeepSeek AI’s innovative step signals a new peak in accessibility and practicality for AI models, potentially fueling the widespread adoption of AI technology across various sectors. _Source: Machine Heart._
【来源】https://www.jiqizhixin.com/articles/2024-05-07-3
Views: 25