AEROSPACE SHANGHAI, Volume. 42, Issue 2, 1(2025)

Research on the Application of DeepSeek in Aerospace Scientific Research and Production

Bin MIN, Lifang LIN, Jian WU*, Chao MA, and Fuding CHEN
Author Affiliations
  • Shanghai Academy of Spaceflight Technology,Shanghai201109,China
  • show less
    Figures & Tables(2)
    • Table 1. Evolution of DeepSeek large-scale model products

      View table
      View in Article

      Table 1. Evolution of DeepSeek large-scale model products

      模型名称发布时间模型类型主要特点
      DeepSeek Coder2023-11-02代码模型专注于代码生成与理解,架构类似Llama
      DeepSeek LLM2023-11-29通用模型通用大语言模型,通过监督微调提升多任务处理能力
      DeepSeek MoE2024-01-09混合专家模型引入MoE架构,提升模型效率
      DeepSeek Math2024-04数学推理模型专攻数学推理,通过分组相对策略优化(Group Relative Policy Optimization,GRPO)强化训练
      DeepSeek V22024-05通用模型采用MLA和MoE架构,支持128 KB长上下文
      DeepSeek V32024-12通用模型基于V2架构扩展,参数量达671亿,进一步优化多任务处理能力
      DeepSeek R12025-01-20推理模型专注逻辑推理与实时问题解决,参数规模与V3一致(671亿)
      Janus-Pro2025-01-28多模态模型具有更好的视觉质量、更丰富的细节以及生成简单文本
    • Table 2. Methods for V3 to save training costs

      View table
      View in Article

      Table 2. Methods for V3 to save training costs

      模型结构模型训练方式针对性GPU优化
      DeepSeek MoE+MLADual Pipe(双向流水线并行)低精度FP8 (8位浮点数)训练

      无需辅助损失的负载

      均衡

      All To All(全互连)通信内核

      IB(InfiniBand,无限带宽)+ NVLink(NVIDIA推出的一种高速芯片间互联技术)

      PTX(Parallel Thread eXecution,并行线程执行) 语言
      MTP无张量并行TP(Token预测)带宽限制
    Tools

    Get Citation

    Copy Citation Text

    Bin MIN, Lifang LIN, Jian WU, Chao MA, Fuding CHEN. Research on the Application of DeepSeek in Aerospace Scientific Research and Production[J]. AEROSPACE SHANGHAI, 2025, 42(2): 1

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category: Research and Application of Large Models

    Received: Mar. 12, 2025

    Accepted: --

    Published Online: May. 26, 2025

    The Author Email:

    DOI:10.19328/j.cnki.2096-8655.2025.02.001

    Topics