《DeepSeek入门宝典》第1册.技术解析篇 (51CTO)（Z-Library）

【One-Line Pitch】 A compact technical primer for anyone who wants to understand what DeepSeek R1 is, how it was built, and why it matters—without wading through dense research papers. Ideal for developers, AI enthusiasts, and tech decision-makers curious about open-source reasoning models. 【Book Arc】 - **Opening (~0%–10%)**: Introduces DeepSeek as a company and product—founded by Liang Wenfeng in 2023, first model released in January 2024, and the R1 model that went viral in January 2025. Covers the three headline traits: high performance, open-source, and low cost. - **Early (~10%–30%)**: Compares five ways to use DeepSeek—official website, mobile app, API, local deployment via tools like Ollama/vLLM/MNN, and cloud platforms like SiliconFlow, Tencent Cloud, and Alibaba Cloud. Sets up the audience split between casual users and developers. - **Middle (~30%–60%)**: Dives into the technical foundation—DeepSeek V3 as the base MoE model (671B parameters, ~37B active per token, pretrained on 14.8T tokens)—and explains R1's three variants: R1-Zero (pure RL), R1 (multi-stage training), and R1-Distill (knowledge distillation). - **Late (~60%–85%)**: Walks through R1's training pipeline: reinforcement learning with multiple reward signals, cold-start data from long chain-of-thought traces, two stages of supervised fine-tuning, and distillation from the large teacher model to smaller student models. Highlights three key contributions: proving pure RL works, the "aha moment" in reasoning, and distilled small models beating OpenAI o1-mini. - **Ending (~85%–100%)**: Compares R1 with OpenAI o1 across architecture, training method, and ecosystem; outlines four future directions (general capability, language mixing, prompt engineering, software engineering tasks); and closes with a product family overview plus resource links. 【Key Takeaways】 - **DeepSeek R1 is a reasoning-focused model built on top of V3** (Early): V3 is a Mixture-of-Experts model with 671B total parameters but only ~37B active per token, making it computationally efficient. R1 inherits this base and specializes in chain-of-thought reasoning. - **Three traits define R1's appeal: performance, openness, and cost** (Early): It matches OpenAI o1 on reasoning benchmarks, is fully open-sourced with training details public, and reportedly cost only ~2% of o1's development budget. This combination is what drove its explosive adoption. - **Five access paths serve different user types** (Early): Casual users can use the website or app; developers can call the API; enterprises needing stability can deploy locally with Ollama, vLLM, or MNN; cloud platforms offer managed alternatives. Choose based on your need for control versus convenience. - **R1's training uses reinforcement learning as a "scoring system"** (Middle): Multiple reward signals evaluate the model's outputs, guiding it toward better reasoning. This is the core mechanism that drives R1-Zero's improvement without supervised fine-tuning. - **Cold-start data acts as a "tutorial" before main training** (Middle): A small set of high-quality chain-of-thought examples helps the model learn clear, logical reasoning patterns faster. This strategic seeding is what makes the RL phase more effective. - **Two stages of supervised fine-tuning (SFT) refine task performance** (Middle): The model learns from labeled data to become more precise on specific tasks. This complements the RL phase and is a standard but crucial step in the pipeline. - **Distillation transfers reasoning skills to smaller models** (Middle): The large R1 model acts as a teacher, and six smaller distilled models were released to the community. These small models outperform OpenAI o1-mini, proving that distillation is both economical and effective—though pushing the frontier further may still require larger base models and more RL. - **R1 differs from OpenAI o1 in three structural ways** (Ending): Different architecture (built on V3 vs. a new model), different training approach (pure RL vs. RL+SFT combined), and different ecosystem (open-source and free vs. closed and subscription-based). These differences shape how each model can be used and extended. 【Reading Tips】 - **Skim the opening sections** (~0%–30%) if you already know what DeepSeek is—the five usage methods are the only part worth a careful look, especially the local deployment options. - **Deep-read the middle section** (~30%–60%) on V3 and R1's variants. The distinction between R1-Zero, R1, and R1-Distill is the conceptual core of the book; get this clear and the rest falls into place. - **Pay attention to the training pipeline details** (~60%–85%): RL rewards, cold-start data, SFT stages, and distillation. These four concepts explain *why* R1 works and are directly transferable to understanding other LLM training approaches. - **The ending comparison with OpenAI o1** (~85%–100%) is the most practical part for decision-making—it clarifies what you gain and lose by choosing R1 over o1 in real projects. - **Don't worry about the figures and charts**—they're mostly benchmark comparisons from official sources. The text explanations carry the substance; treat the charts as supporting evidence. 【Coverage Limits】 Excerpts cover the technical explanation of DeepSeek R1's architecture, training, and contributions, plus usage methods and future directions. They do not cover hands-on coding examples, prompt engineering tutorials, or detailed API documentation.