DeepSeek V4 Thinking Mode: The Complete Guide — AI Development Tutorial

What is Thinking Mode

The DeepSeek V4 series (Flash and Pro) both support Thinking Mode, a major breakthrough in DeepSeek's reasoning capabilities. When Thinking Mode is enabled, the model performs an internal chain-of-thought analysis before generating the final answer, similar to the Chain-of-Thought mechanism in OpenAI o1/o3, but DeepSeek offers it as an optional feature—a single model supports both non-thinking and thinking modes, easily switched via parameters.

Enabling Thinking Mode

Controlled via the thinking parameter:

from openai import OpenAI
import os

client = OpenAI(
    api_key=os.environ.get('DEEPSEEK_API_KEY'),
    base_url='https://api.deepseek.com'
)

response = client.chat.completions.create(
    model='deepseek-v4-pro',
    messages=[
        {'role': 'system', 'content': 'You are a helpful assistant'},
        {'role': 'user', 'content': 'Please prove that the square root of 2 is irrational'}
    ],
    extra_body={'thinking': {'type': 'enabled'}},
    reasoning_effort='high'
)

print('Thinking process:', response.choices[0].message.reasoning_content)
print('Final answer:', response.choices[0].message.content)

Detailed Explanation of reasoning_effort Levels

Level	Effect	Token Consumption	Use Cases
low	Fast reasoning, lightweight thinking	Minimal	Simple math problems, common sense judgments
medium	Default level, balances speed and depth	Moderate	Moderate complexity reasoning, code debugging
high	Deep reasoning, most complete chain of thought	Maximum	Mathematical proofs, complex logic, competition problems

Extracting and Displaying the Chain of Thought

The thinking content is located in the reasoning_content field. For frontend display, it is recommended to collapse it into an expandable area:

// Node.js example
import OpenAI from 'openai';

const openai = new OpenAI({
  baseURL: 'https://api.deepseek.com',
  apiKey: process.env.DEEPSEEK_API_KEY,
});

const completion = await openai.chat.completions.create({
  messages: [
    { role: 'system', content: 'You are a math expert' },
    { role: 'user', content: 'Solve the equation x2 - 5x + 6 = 0' }
  ],
  model: 'deepseek-v4-flash',
  thinking: { type: 'enabled' },
  reasoning_effort: 'medium',
});

console.log('Reasoning chain:', completion.choices[0].message.reasoning_content);
console.log('Answer:', completion.choices[0].message.content);

Dual-Mode Switching Strategy

The best practice is to dynamically choose the mode based on the task type:

Non-thinking mode: Daily conversation, translation, summarization, simple Q&A—faster and cheaper
Thinking mode: Mathematics, programming, logical reasoning, complex analysis—quality first
Strategy: Use non-thinking mode for quick responses first; if the answer quality is insufficient, automatically upgrade to thinking mode and retry

Pricing Rules

Tokens generated in thinking mode are divided into two categories: reasoning tokens (reasoning_content) are billed at input prices, and output tokens (content) are billed at output prices. The reasoning token price for V4 Flash thinking mode is ¥1/million tokens, and for V4 Pro it is ¥3/million tokens.

Notes

FIM completion only supports non-thinking mode
reasoning_content in thinking mode needs to be parsed separately
In streaming output, thinking content is also returned in chunks
It is recommended to set an appropriate reasoning_effort based on task complexity to avoid waste