近日，DeepSeek AI公司推出了一款新的开源混合专家（MoE）语言模型——DeepSeek-V2，其表现引起了业界广泛关注。该模型主打训练成本更低、推理更加高效，是探索通用人工智能（AGI）领域的重要成果。
DeepSeek-V2参数量达到了惊人的236B，支持长达128K的token上下文长度。其在不同的基准测试中均表现出色，不仅在AlignBench上超越了GPT-4，更逼近了GPT-4-Turbo的表现。此外，在MT-Bench测试中，DeepSeek-V2与LLaMA3-70B相媲美，并优于Mixtral 8x22B。
值得一提的是，DeepSeek-V2在数学、代码和推理方面表现出众，展现了强大的语言理解和应用能力。该模型的开源将进一步推动自然语言处理领域的发展，并有助于降低开发成本和提高效率。
DeepSeek AI公司的这款新模型引起了业界的高度关注，并有望为未来的通用人工智能发展打下坚实基础。目前，DeepSeek-V2已开源供公众使用，相信会有更多的研究者和开发者加入这一领域，共同推动人工智能的发展。
英语如下：
News Title: “DeepSeek Open Source MoE Model Stuns: Approaching GPT-4, a New Benchmark in Efficient General AI”
Keywords: DeepSeek Open Source MoE Model, AI Language Model, Impressive Performance
News Content:
DeepSeek AI Company recently unveiled its new open-source Mixture-of-Experts (MoE) language model, DeepSeek-V2, which has garnered widespread attention in the industry for its impressive performance. The model is touted as a significant achievement in exploring General Artificial Intelligence (AGI) due to its lower training costs and more efficient reasoning.
With an astonishing parameter count of 236B, DeepSeek-V2 supports a contextual token length of up to 128K. It has excelled in various benchmark tests, surpassing GPT-4 on the AlignBench and even approaching the performance of GPT-4-Turbo. Additionally, in the MT-Bench test, DeepSeek-V2 is on par with LLaMA3-70B and superior to Mixtral 8x22B.
Notably, DeepSeek-V2 stands out in mathematics, coding, and reasoning, demonstrating powerful language understanding and application capabilities. The model’s open-source status will further propel the development of natural language processing and contribute to reducing development costs and improving efficiency.
This new model from DeepSeek AI has garnered significant attention in the industry and is expected to solidify the foundation for future advancements in general AI. Currently, DeepSeek-V2 is open-source and available for public use, inviting more researchers and developers to join the field and drive the advancement of artificial intelligence together.
【来源】https://www.jiqizhixin.com/articles/2024-05-07-3
Views: 1