DeepSeek 图解 10 页 PDF 作者：郭震 2025.2.3 目录 1 本地部署并运行 DeepSeek . . . . . . . . . . . . . . . . . . . . . . 2 1.1 为什么要在本地部署 DeepSeek . . . . . . . . . . . . . . . . . 2 1.2 DeepSeek 本地部署三个步骤 . . . . . 的中间推理模型训练过程 . . . . . . . . . . . . . . 9 3.3 通用强化学习训练过程 . . . . . . . . . . . . . . . . . . . . . . 10 3.4 总结 DeepSeek-R1 . . . . . . . . . . . . . . . . . . . . . . . . 11 4 参考文献 . . . . . . . . . . 料用心打磨且开源，是为了帮助更多人了解获取 AI 知识，严禁拿此资料引流、出书、等形式的商业活动 图 1: 我的公众号：郭震 AI 安装后，打开命令窗口，输入 ollama，然后就能看到它的相关指令，一共 10 个左右的命令，如下图2所示，就能帮我们管理好不同大模型： 图 2: Ollama 常用的命令 第二步，命令窗口输入：ollama pull deepseek-r1:1.5b，下载大模型 deepseek-

可解释性：注重模型输出 的可解释性和透明性。 DeepSeek R1  高效推理：专注于低延迟和 高吞吐量，适合实时应用。  轻量化设计：模型结构优化， 资源占用少，适合边缘设备 和移动端。  多任务支持：支持多种任务， 如文本生成、分类和问答。 Kimi k1.5  垂直领域优化：针对特定领域 （如医疗、法律）进行优化， 提供高精度结果。  长文本处理：擅长处理长文本 alloy-based anodes such as silicon (Si, 4200 mA h g-1) show extremely high theoretical capacity, nearly 10 times higher than the capacity of commercial graphite anodes (372 mA h g-1). Unfortunately, these and discussion. This is primarily due to their extremely high theoretical capacity, which is nearly 10 times that of commercial graphite anodes (372 mA h g-1). Despite their potential, these alloy-based

演化、动态自适应、动态识别、人机协同感知、人机协同决策与 控制等标准。 9. 智能体标准。规范以通用大模型为核心的智能体实例和 10 智能体基本功能、应用架构等技术要求，包括智能体强化学习、 多任务分解、推理、提示词工程，智能体数据接口和参数范围， 人机协作、智能体自主操作、多智能体分布式一致性等标准。 10. 群体智能标准。规范群体智能算法的控制、编队、感知、 规划、决策、通信等技术要求和评测方法，包括自主控制、协同 载工具、 智能移动终端、数字人、智能服务等标准。 1. 智能机器人标准。规范人工智能在机器人领域应用的技 术要求，包括机器人智能认知、智能决策等标准。 2. 智能运载工具标准。规范智能运载工具感知、识别与预 判、协同与博弈、决策与控制、评价等技术要求，包括环境融合 感知、智能识别预判、智能决策控制、多模式测试评价等标准。 3. 智能移动终端标准。规范人工智能应用在移动终端领域 的技 的技术要求，包括图像识别、人脸识别、智能语音交互，以及智 11 能移动终端涉及的信息无障碍、适老化等标准。 4. 数字人标准。规范数字人的外形、动作生成、语音识别 与合成、自然语言交互等技术要求，包括数字人基础能力评估、 多媒体合成渲染、基础数据采集方法、标识和识别方法等标准。 5. 智能服务标准。规范基于大模型、自然语言处理、智能 语音、计算机视觉等人工智能技术提供的服务，包括模型即服务

engineering 7 LLM output configuration 8 Output length 8 Sampling controls 9 Temperature 9 Top-K and top-P 10 Putting it all together 11 Prompting techniques 13 General prompting / zero shot 13 One-shot probabilities are then sampled to determine what the next produced token will be. Temperature, top-K, and top-P are the most common configuration settings that determine how predicted token probabilities diverse or unexpected results. A temperature of 0 (greedy decoding) is Prompt Engineering February 2025 10 deterministic: the highest probability token is always selected (though note that if two tokens have

Number of Developers, MM 0% 50% 100% Internet LLM 33 Years In 90% @ Year 3 90% @ Year 23 10/22 4/25 800MM Big Six* USA Technology Company CapEx *Apple, NVIDIA, Microsoft, Alphabet, Amazon six leading global LLMs. Source: YipitData (5/25) Desktop User Share, % 2/24 2/25 4/25 75% 60% 10% 21% 15% 0% Details on Page 293 USA – LLM #1 China USA – LLM #2 AI Model Compute Costs High / some, the evolution of AI will create a race to the bottom; for others, it will create a race to the top. The speculative and frenetic forces of capitalism and creative destruction are tectonic. It’s undeniable

show that, even with only 21B activated parameters, DeepSeek-V2 and its chat versions still achieve top-tier performance among open-source models. The model checkpoints are available at h t t p s : / / g . . . . . 9 2.2.3 Auxiliary Loss for Load Balance . . . . . . . . . . . . . . . . . . . . . . . . 10 2.2.4 Token-Dropping Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 3 Pre-Training DeepSeek-V2 still achieves top-tier performance among open-source models and becomes the strongest open-source MoE language model. Figure 1(a) highlights that, on MMLU, DeepSeek-V2 achieves top-ranking performance

to invest 
 in this new infrastructure in ways that will help 
 us grow revenue. Chris Hyams CEO 10 AI in the EnterpriseLesson 3 Start now and invest early How Klarna benefits from AI knowledge compounding automation platform. It works on top of our existing workflows and systems to automate rote work and accelerate insight and action. Our first use case: working on top of Gmail to craft customer responses full ownership. Enterprise-grade compliance Data is encrypted in transit and at rest, aligned with top standards like SOC 2 Type 2 and CSA STAR Level 1. Granular access controls You choose who can see

[和| Alibaba AL.Labs 阿里巴巴人工智能实验室 HIFI 4 Alibaba Al.Labs 阿里巴巴人工智能实验室 Resolution 1. GEMM Tensorize (10x speed up) 2. HIFI4 Program (don't need dlopen) Serial Communication HIFI4 DSP including what you want to compute and how to compute. w What you want to compute @autotvm.register top compute(conv2d，pvr, [direct]) def conv2d_pvr(cfg, data, kernel, strides, padding, dilation, layout

. . . . . . . . . . . . 10 2.4.4 ChatDev . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 2.4.5 solo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 2.4.6 Cursor . . . . . . . . . . . . . . . . . . . 10 2.4.7 Tabby . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 2.4.8 Codeium . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 2.4.9 GitHub Copilot . . . . . . . . . . . . . . . . . . . . . . . . 10 2.4.10 通义灵码 . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 2.5 AI 指令编写工具 . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 2.5.1 FlowGPT

3 Multi-Vendor Support MOTIVATION©2019 HARMAN INTERNATIONAL INDUSTRIES, INCORPORATED 4 Terasic DE10-Nano DEPLOY VTA ON INTEL FPGA©2019 HARMAN INTERNATIONAL INDUSTRIES, INCORPORATED 5 Software - CMA Setup Environment Variables Navigate to 3rdparty/cma and build kernel module Copy kernel module to DE10-Nano and Install Module CMA API Reference©2019 HARMAN INTERNATIONAL INDUSTRIES, INCORPORATED 7 Software INTEL FPGA©2019 HARMAN INTERNATIONAL INDUSTRIES, INCORPORATED 8 Hardware Configure Chisel VTA for DE10-Nano DEPLOY VTA ON INTEL FPGA©2019 HARMAN INTERNATIONAL INDUSTRIES, INCORPORATED 9 Hardware Datapath