DeepSeek Model Complete Guide

The official DeepSeek GitHub repository is at github.com/deepseek-ai, and three core model series have been released:

DeepSeek V3 — 671B parameter MoE (Mixture of Experts) architecture, activating only 37B parameters per inference. Ranked first among open-source models, with performance comparable to GPT-4. Model size is 404GB, supporting a 160K context window.

DeepSeek R1 — A reasoning-enhanced model trained with reinforcement learning. It excels in math, coding, and logical reasoning tasks, with performance close to OpenAI O3 and Gemini 2.5 Pro. The latest version is DeepSeek-R1-0528. Downloads on Ollama exceed 88.3M+, available in 7 sizes from 1.5B to 671B.

DeepSeek Coder — A code-specialized model trained on 2 trillion tokens (87% code + 13% Chinese and English natural language). Ollama downloads exceed 4.2M+, available in 1.3B / 6.7B / 33B sizes.

DeepSeek V3 Architecture

DeepSeek V3 adopts the MoE (Mixture of Experts) architecture, with a total of 671B parameters, activating only 37B parameters per token during inference. This significantly reduces inference cost while maintaining extremely high model capacity. Core technologies include:

DeepSeek R1 Architecture

DeepSeek R1 is a reasoning-enhanced model based on reinforcement learning, achieving deep reasoning capabilities through Chain-of-Thought technology. The R1-0528 update significantly improves reasoning performance. Core technologies include:

R1-0528 Performance Comparison

Step 1: Install Ollama

Visit ollama.com/download and select the installer for your operating system. Supports Windows, macOS, and Linux.

# macOS / Linux can also install via command line
curl -fsSL https://ollama.com/install.sh | sh

Step 2: Download and Run DeepSeek Models

After installation, run the following command in the terminal to download and run the model. The first run will automatically download, and subsequent runs will start directly.

# Recommended for beginners: DeepSeek R1 8B (5.2GB, 128K context)
ollama run deepseek-r1:8b

# Lightweight: DeepSeek R1 1.5B (1.1GB, runs on CPU)
ollama run deepseek-r1:1.5b

# Coding-focused: DeepSeek Coder 6.7B (3.8GB)
ollama run deepseek-coder:6.7b

# Full version: DeepSeek V3 671B (404GB, requires multi-GPU cluster)
ollama run deepseek-v3

Step 3: Start Chatting

Once downloaded, the terminal will display the chat interface directly. Type your questions to interact with DeepSeek. Press Ctrl+D to exit.

API Access

After Ollama starts, it automatically provides an API service at localhost:11434, compatible with OpenAI format:

# curl call
curl http://localhost:11434/api/chat -d '{
"model": "deepseek-r1:8b",
"messages": [{"role": "user", "content": "Hello, introduce yourself"}]
}'

# Python call (pip install ollama)
from ollama import chat
response = chat(
model='deepseek-r1:8b',
messages=[{'role': 'user', 'content': 'Hello'}]
)
print(response.message.content)

# JavaScript call (npm install ollama)
import ollama from 'ollama'
const response = await ollama.chat({
model: 'deepseek-r1:8b',
messages: [{role: 'user', content: 'Hello'}]
})
console.log(response.message.content)

Advanced: Integrate with Development Tools

Ollama supports integrating DeepSeek R1 into development tools like Claude Code, Codex, etc.:

# Use DeepSeek R1 in Claude Code
ollama launch claude --model deepseek-r1:8b

# Use in Codex App
ollama launch codex-app --model deepseek-r1:8b

# Use in OpenCode
ollama launch opencode --model deepseek-r1:8b