Peking University's DeepSeek Internal Seminar Series (PDF file) - AI Tutorial Materials

The "DeepSeek Internal Seminar Series" launched by Peking University mainly introduces DeepSeek.artificialintelligentModels and applications.DeepSeek Principles and Practical ApplicationsThis section elucidates the principles, development history, and technological breakthroughs of the DeepSeek model, including its reasoning capabilities,MultimodalSupport andopen sourceThe presentation highlights DeepSeek's key advantages. It also showcases DeepSeek's applications in education, academia, and paid knowledge platforms, including instructional design, homework grading, and paper writing assistance, emphasizing its low cost and high performance. "DeepSeek Private Deployment and Appliances" focuses on DeepSeek's private deployment, including deployment methods for individual users on local computers, enterprise-level high-performance deployment solutions, and low-cost deployment strategies for scenarios with limited computing resources. It also introduces the configuration, performance data, and application scenarios of DeepSeek appliances, providing a comprehensive solution for different users, from model selection to practical implementation.

Get theDeepSeek Principles and Practical Applications》And DeepSeek Private Deployment and Appliances

Original PDF fileScan the QR code to follow and reply: 20250225

DeepSeek Principles and Practical Applications

"DeepSeek Principles and Practical Applications" is a book published by Peking University Qingniao.artificialintelligentThis internal handbook, jointly published by the Research Institute, the Metaverse Technology Institute of the School of Computer Science, and the Learning Science Laboratory of the School of Education, details the principles of DeepSeek technology and its applications across various fields. It showcases DeepSeek's wide-ranging applications in education, such as instructional design, teaching activities, homework grading, personalized lesson plans, medical pathology diagnosis, full-process thesis assistance, academic research, paid knowledge services, and homework tutoring, providing specific examples.Prompt wordsExamples are provided to help users better leverage DeepSeek to improve work efficiency and learning outcomes. The application potential of DeepSeek in reasoning-intensive tasks, document analysis and long-context understanding, open-domain question answering and writing, etc., is explored.AIThe widespread adoption of technology and the flourishing ecosystem of domestic production have provided strong support.

artificialintelligentConceptual Analysis

• Large ModelRelated terms:introduceMultimodal(Text, images, audio, video)AIThe concepts of tools (such as DeepSeek, Doubao, etc.), general models (large language models, visual models, etc.) and industry models (education, healthcare, finance, etc.).
• Large ModelPast and present lives:fromartificialintelligentStarting from its development history, this section introduces computational reasoning, knowledge engineering, and...Machine LearningArtificialNeural Networks,depthNeural NetworksThe development trajectory of technologies such as Transformer architecture, encoders, and decoders, and generative...artificialintelligent(Gen)AIThis includes content such as Diffusion architecture.

• Large ModelDevelopment stagesCompare the characteristics of generative models and inference models, such as Open...AI GPT-4o (Generative Model) and OpenAI o1 (reasoning model) in model positioning, reasoning ability,MultimodalDifferences in support, application scenarios, and user interaction experience.
• Generative Models and InferenceLarge ModelComparisonGenerative models are suitable for general tasks andMultimodalInteractive reasoning models perform better in complex logic and specialized tasks, but their interaction pace is slower.MultimodalSupport is limited.

DeepSeek R1

• Company Background and Development HistoryDeepSeek was born from Magic Square Quant and is a company focused onartificialintelligentDeepSeek is an innovative technology company focused on basic technology research and development. This section introduces DeepSeek's founding background, investor background, market positioning, and development history.
• DeepSeek Model SeriesThis comparison examines the design intent, performance, and application scope of DeepSeek's generative model (V3) and inference model (R1). The inference model (R1) performs exceptionally well in complex inference tasks, achieving an F1 score of 92.2% on the DROP task.AIThe pass rate for ME 2024 was 79.8%.
• Why is it hot?Emphasizing the three major advantages of DeepSeek R1—open sourceLow cost, domestic production, and breakthroughs in reasoning capabilities have made it a global leader.Large ModelA key leader in the field.

• DeepSeek's algorithmic and computing power breakthroughsThis section introduces training techniques for DeepSeek R1, such as Hybrid Expert (MOE), Multi-Head Latent Attention (MLA), and Reinforcement Learning (GRPO), and their applications.AISignificant impacts on the industry include breaking monopolies, lowering prices, and driving innovation.
• rightAISignificant impact on the industryDeepSeek-R1 is based on low cost andopen sourceFeatures break monopolies, lower prices, and drive growthAIIndustry innovation.
• How to useThis section introduces three ways to use DeepSeek: calling the official API, model fine-tuning, and direct use, providing multiple access methods such as the official website, API, mobile APP, and third-party channels.
• Private DeploymentThis section introduces private deployment methods for the model, including Ollam deployment and vLLMDeployment, etc., listing different versions of the DeepSeek-R1 model and their characteristics.

• GPTWorking principleDetailed introductionGPTThe working principle includes probabilistic prediction, word chain, Transformer architecture, pre-training, supervised fine-tuning, human feedback reinforcement learning, as well as data sources, model parameters, and context window size.

• Advantages and disadvantages of generative modelsCompare the advantages (such as language understanding and generation capabilities, world knowledge capabilities, and certain reasoning capabilities) and disadvantages (such as hallucinations, limited knowledge base, and context window limitations) of generative models.
• How DeepSeek-R1 worksThis section introduces the reasoning mechanism of the DeepSeek R1 model, including techniques such as Chain of Thought, Distillation, and Reinforcement Learning, and provides links to relevant papers.

• DeepSeek's conversational capabilities: Emphasizing that DeepSeek enables humans to interact withAIThe dialogue is moreSimple,based onPrompt wordsTechniques to improve conversation quality, suitable for users at all levels.fastGet started.

• DeepSeekPrompt wordsSkillList of multiplePrompt wordsTechniques such as sincerity and directness, universal formulas, speaking in plain language, reverse PUA, being good at imitation, being good at sharp commentary, and stimulating deep thinking help users better utilize DeepSeek for task processing.
  • Sincerity + DirectnessClearly define the task objectives, directly express the requirements, and avoid vague expressions.
  • general formulaOrganize according to the formula: "I want to do it, I want it to be used, I hope it will achieve the desired effect, but I am worried about the problems."Prompt wordsThis helps DeepSeek fully understand the task.
  • • Speak human languageDeepSeek is required to answer in plain and easy-to-understand language, avoiding overly technical or official expressions.
    • Reverse PUAAsk challenging questions (such as reasons for opposition or angles of criticism) to stimulate DeepSeek’s in-depth thinking and comprehensive answers.
    • • Good at imitation: Specify the object or style to imitate, and let DeepSeek generate content that matches a specific style or tone.
      • Skilled in sharp commentaryUse specific sentence structures or expressions (such as "laughing to death") to stimulate DeepSeek's critical thinking and sharp commentary abilities.
      • Stimulate deep thinking:existPrompt wordsThe requirements for "review" and "critical thinking" have been added to improve the depth and quality of the answers.
      • DeepSeek OfficialPrompt wordsInterpretation: Introduction to the DeepSeek official documentationPrompt wordsCategories, such as code, content, structured output, role-playing, creative, translation, etc.Prompt wordsIt generates, etc., and provides links.
      • Application scenariosList the application scenarios of DeepSeek R1, including reasoning-intensive tasks (such as programming tasks and solving mathematical problems), education and knowledge applications (such as instructional design, teaching activities, homework correction, personalized lesson plans, medical pathology diagnosis, and full-process paper writing assistance), open-domain question answering and writing, etc.
      • Education and academic empowermentThis section details DeepSeek's applications in education, including instructional design, teaching activities, homework grading, personalized lesson plans, medical pathology diagnosis, full-process paper writing assistance, academic research, paid knowledge services, and homework tutoring, providing specific examples.Prompt wordsExample.
        • Instructional DesignDesign the course outline, clarifying the course content and design concept.
        • Teaching activitiesDesign variations of questions, problem scaffolding, homework assignments, and group practice scripts.
        • • Homework correctionAdd explanations to objective questions and provide ideas for programming questions.
          • Personalized lesson plansDesign sets of exercises of varying difficulty and label them with difficulty levels.
          • Medical Pathology DiagnosisGenerate a list of disease diagnoses, sort them, and suggest examinations.
          • Full-process assistance for thesis writingServices include: assisting with topic selection, generating outlines, literature reviews, expanding content, and proofreading/plagiarism checking.
          • • academic researchServices include: assisting with rapid reading of references, compiling references, academic translation, designing research methods, and writing data analysis reports.
            • Knowledge paymentDesign course outlines, live streaming scripts, community operation scripts, and paid Q&A scripts.
            • Homework tutoringFrom anchoring knowledge points to expanding capabilities, design a tutoring process.

            DeepSeek Private Deployment and Appliances

            "DeepSeek Private Deployment and All-in-One Machine" is a book by Beida Jade Bird.artificialintelligentThis guide, jointly released by the Research Institute and the Metaverse Technology Institute of the School of Computer Science at Peking University, details the private deployment methods for DeepSeek models, including deployment solutions for individual and enterprise users. It covers the entire process from model selection to deployment, providing optimization suggestions for different hardware environments and business needs. A key focus is the DeepSeek appliance, a device integrating high-performance hardware and pre-installed models, providing out-of-the-box functionality for enterprises and educational institutions. AI Solutions. We provide detailed specifications and performance data for all-in-one machines in different configurations to help businesses choose the appropriate model based on their needs.

            artificialintelligentWith DeepSeek

            • Large ModelRelated terms:introduceMultimodal,AITools (such as DeepSeek, Doubao, etc.), general models (including large language models, generative models, inference models, etc.), and industry models (such as models in education, healthcare, finance, etc.).
            • Large ModelPast and present lives:reviewartificialintelligentThe development history includes computational reasoning, knowledge engineering,Machine LearningArtificialNeural Networks,depthNeural NetworksTransformer architecture, and generative architecture, etc.artificialintelligent(Gen)AIThe rise of ) , listing various domestic and international examplesLarge Model.
            • Large ModelDevelopment stagesComparing the localization and reasoning capabilities of generative and inference models.MultimodalSupport, application scenarios, and user interaction experience.
            • Generative Models and InferenceLarge ModelComparison:Generative models excel atMultimodalContent generation and general tasks are suitable for mass applications; the reasoning model performs well in complex logical reasoning and professional fields, making it suitable for high-precision reasoning needs.
            • Starting in 2024,AIXiao Rui's team then integrated DeepSeek V2:introduceAIXiao Rui's team's research and application of the DeepSeek technology system includes the evolution of V2 (MoE architecture), V3 (MTP) and R1 (inference model), as well as the application of knowledge distillation in various versions.
            • DeepSeekfastGoing viralThis article discusses the release of DeepSeek V3 and R1 models and their impact on the industry, as well as the widespread popularity of DeepSeek and the R1 model after the 2025 Spring Festival.
            • Company ProfileThis section introduces DeepSeek's founding background, team composition, model series, and its connections with other Chinese companies.Large ModelRelationships between enterprises.
            • The evolutionary history and characteristics of the modelCompare the core architecture, training methods, key features, and performance of the three main DeepSeek models: V2, V3, and R1.
            • Advantages of the DeepSeek modelSummarize the characteristics of the DeepSeek model, such asopen sourceEcosystem, model distillation support, low-cost training, etc.
            • Summary of DeepSeek Model FeaturesThis section provides a detailed introduction to the basic architecture of DeepSeek models (such as MoE, MLA, RL, etc.), training and frameworks (such as FP8 mixed precision training, TTC, etc.), and their social value.
            • DeepSeek-R1 suiteList the various distillation models of DeepSeek-R1, their origins, and characteristics.
            • Distillation Model - ConceptExplain the definition, principle, advantages, and application of model distillation in DeepSeek.
            • DeepSeek-R1 Distillation Model - Capability ComparisonIt provides experience in selecting distillation models, including applicable scenarios and performance of models with different parameter values.
            • DeepSeek model application scenariosThis section introduces the application advantages of DeepSeek in edge computing, complex mathematical and programming tasks, and Chinese language scenarios.
            • How to useThis article summarizes the various ways to use DeepSeek, including using official services directly, third-party services and API calls, personal deployment, and enterprise deployment.

            Personal deployment of DeepSeek

            • Model Deployment:
              • Hardware configuration -recommendProvides models with different parameter valuesrecommendHardware configuration, including CPU, memory, video memory, and hard disk space requirements.
              • Hardware Configuration - MinimumProvides the minimum hardware configuration requirements for models with different parameter values.
              • Hardware selection recommendationsBased on requirements, budget, and performance balance, hardware selection recommendations are provided.
              • • Ollama FeaturesIntroduce the features of Ollam tools, such as...open sourcefree,SimpleEasy to use, rich in models, etc.
                • Software installation stepsDetailed instructions on installing the Ollam client and Docker.
                • Environment variable configurationThis section explains how to configure Ollam environment variables and provides suggestions.
                • Network configurationThis section explains how to access the Ollam service, including network configuration considerations for local and remote deployments.
                • Environmental preparation:
                  • Ollama Common CommandsList commonly used Ollam commands, such as downloading models, running models, and viewing model information.
                  • Download the model using the Ollam command.This section introduces the command format for the download model, the monitoring of the download process, and the methods for handling download failures.
                  • Run the model using the Ollam command.This section introduces the command format for running the model, the methods for checking the running status, and how to stop running the model.
                  • Model running effect: Show the effect after the model is run.
                  • Ollama FAQsThis section discusses solutions to common security vulnerabilities, unauthorized access risks, and performance degradation issues after model upgrades in Ollam.
                  • Front-end display:
                    • Front-end displayThis section introduces the features, applicable scenarios, and usage methods of front-end presentation tools such as PageAssist, Chatbox, and OpenWebUI.
                    • Summary of the three interfacesCompare the features and applicable scenarios of PageAssist, Chatbox, and OpenWebUI.

                    Enterprise deployment of DeepSeek

                    • Introduction to TransformersThis section introduces the features and functionalities of the Transformers library, including supported modal tasks and framework interoperability.
                    • Transformers Deployment ModelThis section details the steps for deploying a DeepSeek model using Transformers, including installing dependencies, loading the model, running the model, and making other framework calls.
                    • Commonly used parameters during generationList the commonly used parameters, their meanings, and precautions when generating the model.
                    • vLLMIntroductionIntroduction vLLMThe framework features include PagedAttention technology, performance improvements, and seamless integration with Transformers.
                    • vLLMDeployment Model: Explanation of using vLLMThe steps to deploy a DeepSeek model include installing dependencies, loading and starting the model.
                    • vLLMParameter Description: Detailed explanation of vLLMThe parameters in the startup command and their functions.
                    • vLLMActual concurrent performance testingProvide vLLMConcurrency performance test data under different device and model configurations.
                    • vLLMCommon problems and solutionsDiscussion vLLMAttention mechanism compatibility issues that may be encountered during deployment and their solutions.
                    • Enterprise-level 70B model deployment case sharingThis presentation will share enterprise-level 70B model deployment case studies, including server configuration, deployment solutions, and experience summaries.
                    • Deployment Solution OverviewThis section introduces the deployment scheme for the full-fledged DeepSeek R1 model, including hardware requirements and cost analysis.
                    • Solution ImplementationThis document details the hardware environment configuration, model download, and implementation steps for three low-cost deployment schemes (llama.cpp, KTransformers, and Unsloth dynamic quantization).
                    • Option 1, llama.cppThis section introduces the background, installation steps, reasoning steps, and test results of the llama.cpp project.
                    • Option 2, KTransformersThis section introduces the background, installation steps, model running methods, and practical results of the KTransformers project under single-concurrency conditions.
                    • Option 3: Unsloth dynamic quantization + OllamThis section introduces how to merge the weight files of the Unsloth dynamic quantization model using the llama.cpp tool, and how to register and call the model using the Ollam tool.
                    • Performance Evaluation SummaryCompare the performance of three low-cost deployment schemes under different hardware configurations.
                    • Deployment Plan SummaryThis paper summarizes the applicability, advantages, and limitations of three low-cost deployment options.

                    DeepSeek All-in-One Machine

                    • DeepSeek All-in-One MachineThis section introduces the concept, advantages, and applicable scenarios of the DeepSeek all-in-one machine.
                    • List of domestic DeepSeek all-in-one machine manufacturersList the domestic DeepSeek all-in-one machine manufacturers and their product features as of February 2025.
                    • DeepSeek 671B All-in-One PC (Chinese version)recommendConfigurationProvides hardware specifications and applicable scenarios for DeepSeek all-in-one machines with different configurations.
                    • DeepSeek All-in-One Machine Data SharingThis section shares performance data of the DeepSeek appliance under different configurations, including concurrency capabilities and the number of users supported.
                    • Beida Jade BirdAILaboratory Case SharingIntroduction to the services provided by Beida Jade Bird to universitiesAILaboratory construction plan, including the configuration and application scenarios of the DeepSeek all-in-one machine.
                    • Beida Jade BirdAILaboratory construction planDetailed ExplanationAIThe characteristics and functions of the system layer, container layer, business layer, monitoring layer and application layer of the laboratory construction plan.
                    • Beida Jade BirdAILaboratory (DeepSeek All-in-One Machine)This section provides detailed configurations, performance, and pricing information for the DeepSeek Basic, Standard, and Flagship all-in-one printers.

                    Get"DeepSeek Internal Seminar Series" and "DeepSeek Private Deployment and Appliances"

                    Original PDF fileScan the QR code to follow and reply: 20250225

                    China in 2024artificialintelligentIndustry Research Report (PDF file)

                    Zhejiang University DeepSeek Series Lecture Courseware, Session 4 (PDF file) AITutorial materials