What is DeepSeek R1? DeepSeek R1 Tutorial

DeepSeek is a large language model (LLM) that competes with international AI leaders such as OpenAI and Anthropic. However, DeepSeek focuses more on achieving AGI, making AI more general and intelligent.

Compared to OpenAI’s closed strategy, DeepSeek is built around open-source, low-cost, and high-efficiency principles, attracting global developers and enterprises. Its core technologies include:

• Multi-head Latent Attention (MLA): Reduces memory usage and accelerates long-text processing through low-rank factorization technology.
• Mixture of Experts (MoE) architecture: Activates only part of the parameters for computation, improving efficiency and reducing energy consumption.
• FP8 mixed-precision training: Compared to traditional FP16 and FP32, it saves more memory and enhances computational speed.
• DualPipe communication optimization: Improves data transmission between GPUs and reduces latency.

What is AGI?

AGI stands for Artificial General Intelligence. It is still under development, with the goal of making AI as intelligent as humans, capable of understanding and learning, diversifying its functions, thinking like a human, and even possessing self-awareness.

DeepSeek R1 is an advanced AI model based on the DeepSeek-V3 architecture, focusing on mathematical reasoning, code generation, and logical inference. DeepSeek R1 uses pure reinforcement learning (RL) for training, skipping supervised fine-tuning (SFT), reducing manual annotation costs, and improving reasoning ability and generalization.

Core Technical Features

Key features include:

• Reinforcement learning training: Discards SFT, allowing AI to explore problem-solving strategies independently.
• Cold start mechanism: Fine-tunes the model with a small number of high-quality "Chain of Thought" (CoT) examples to enhance reasoning.
• Multi-stage training strategy: Combines mathematical, programming, logical, and scientific problems to improve model consistency by avoiding language interference.
• Mixture of Experts (MoE) architecture: Contains 671 billion parameters but only activates 37 billion per run, balancing performance and computational resources.
• Dual reward mechanism: Ensures answer accuracy and completeness of the reasoning process through "correctness rewards" and "format rewards."

DeepSeek R1 VS. DeepSeek V3

DeepSeek R1 VS. GPT-o1

Mobile and Web Versions

• iOS: Download "DeepSeek" from the App Store
• Android: Available on Google Play or official APK
• Web: Search for DeepSeek Web Version

Local Offline Use

Installing Ollama

DeepSeek R1 relies on Ollama for local execution. First, install Ollama:

• Search for the official website: Ollama
• Download the version for your OS (Windows/macOS/Linux)

Downloading DeepSeek R1

DeepSeek R1 offers multiple parameter sizes (1.5B, 7B, 8B, 14B, 32B, 70B, 671B). Choose one based on your computer's specifications. Example for 7B:

ollama run deepseek-r1:7b

Verify Installation

After downloading, check the installation with:

ollama list

Running DeepSeek R1

Launch the model using:

ollama run deepseek-r1

After completing the above steps, you can start using DeepSeek R1!

As a leading AI product in China, DeepSeek R1 has gained significant attention for its reasoning capabilities and open-source strategy. However, it has also sparked controversy regarding its technology sources, data privacy, and content censorship. Here are the key issues:

Technology Source and Intellectual Property Dispute

Core Issue: Has DeepSeek violated OpenAI's Terms of Service?

• Data distillation concerns: According to Reuters, DeepSeek may have used data distillation techniques, leveraging OpenAI’s outputs to train its models. While this can accelerate AI improvements, it may violate OpenAI’s terms.
• OpenAI's Response: OpenAI has not yet taken legal action, but industry experts worry that data distillation could harm fair competition and lead to stricter AI regulations.
• DeepSeek's Position: DeepSeek denies any violations, emphasizing that its models are independently trained and not based on OpenAI’s outputs.

Content Censorship and Freedom of Speech

Core Issue: Is DeepSeek R1 controlled by the Chinese government?

• Censorship of sensitive topics: DeepSeek restricts discussions on sensitive topics related to the Chinese government.
• Geopolitical bias: DeepSeek R1 may provide more conservative responses regarding China's relations with other countries.

Data Privacy and Security Risks

Core Issue: Does DeepSeek leak user data to the Chinese government?

• Data collection concerns and government regulations: Under China’s National Security Law, tech companies are required to provide necessary data to the government. DeepSeek may collect user chat logs, IP addresses, and keystroke patterns, which could be used for monitoring or data analysis.

Want to explore more AI tools after learning about DeepSeek R1? Try GenApe AI Tool