DeepSeek V3 Open Source China AI 2025.12

685 billion parameter MoE architecture, $5.57 million training cost comparable to Claude 3.5

685B Total parameters
37B Activation parameters
$5.57M Training cost
MIT Open Source Agreement

🔥 Latest update in December 2025

DeepSeek V3-0324 Release: On December 26, 2025, Magic Square Quantization officially released the latest version of DeepSeek-V3. The new version has greatly improved its programming capabilities, and the code generation quality is close to that of Claude 3.7 Sonnet. The open source version has been launched on Hugging Face, with a model volume of 685 billion parameters and a looser MIT open source license.
December 26, 2025

The latest version of DeepSeek V3 is released, greatly improving programming capabilities

March 25, 2025

DeepSeek V3-0324 version released, catching up with Claude 3.7

January 2025

DeepSeek-R1 is released, with reasoning capabilities comparable to OpenAI o1

December 2024

The first version of DeepSeek V3 was released, shocking the industry

🐋 What is DeepSeek V3?

DeepSeek V3 is an open source large language model developed by Deep Search Company, a subsidiary of Hangzhou Magic Square Quantification. It adopts MoE (Mixed Expert) architecture with a total parameter volume of 685 billion, but only 37 billion parameters are activated in a single inference, achieving extremely high cost performance. The most shocking thing is that the training cost of the entire model is only $5.576 million, but it achieves results comparable to Claude 3.5 Sonnet.

🏗️ MoE Architecture

Hybrid expert model, only 37B of 685B parameters are activated, efficient reasoning

💰 Ultra low cost

Training cost of US$5.57 million, only 1/10 of the same level model

📖 MIT Open Source

Open source under MIT license, commercially available, unlimited use

💻 Programming enhancement

The programming capabilities of the latest version have been greatly improved, close to Claude 3.7

🚀 DeepSeek Model Series

DeepSeek-V3 (latest version) 2025.12

The latest released V3 version, based on improved training on a 32K GPU cluster, greatly improves programming capabilities.

685B
Total parameters
37B
Activation parameters
128K
Context
MIT
Open Source Agreement

DeepSeek-R1 Inference model

A model focusing on reasoning capabilities. Through large-scale reinforcement learning training, the reasoning capabilities are comparable to OpenAI o1.

Parity o1
Reasoning ability
RL training
Reinforcement Learning
Thought chain
Reasoning process
Open Source
Agreement

DeepSeek-Coder Code only

A model optimized specifically for code generation and understanding, supporting multiple programming languages.

Code Specialization
Features
Multilingual
Programming Language
FIM
fill mode
Open Source
Agreement

💻 API call example

DeepSeek API is compatible with OpenAI format and can directly use OpenAI SDK:

from openai import OpenAI

client = OpenAI(
    api_key="your_deepseek_api_key",
    base_url="https://api.deepseek.com"
)

response = client.chat.completions.create(
    model="deepseek-chat",  # 或 deepseek-reasoner
    messages=[
        {"role": "system", "content": "你是一个专业的编程助手"},
        {"role": "user", "content": "用 Python 实现快速排序算法"}
    ],
    temperature=0.7,
    stream=True
)

for chunk in response:
    print(chunk.choices[0].delta.content, end="")

💰 Price comparison (per million tokens)

Model Enter price Output price Compare
DeepSeek V3 ¥1 ¥2 Baseline
DeepSeek R1 ¥4 ¥16 4-8x
Claude 3.5 Sonnet ¥21 ¥105 21-52x
GPT-4o ¥18 ¥54 18-27x
💡 King of cost performance: DeepSeek V3's API price is only about 2-5% of Claude 3.5 Sonnet, but performs comparably on multiple benchmarks. This makes it ideal for large-scale AI applications.

🎯 Typical application scenarios

💻 Code generation

The latest version has greatly improved programming capabilities and is suitable for AI programming assistants

🔍 Data Analysis

Ultra-low cost processing of large-scale data analysis tasks

📝 Content Creation

High-quality text generation, supporting a variety of creative scenarios

🤖 Private deployment

MIT open source protocol, can be freely deployed within the enterprise