DeepSeek V3 vs R1: Staged Training vs Iterative SFT-RL Cycles

The AI landscape has been revolutionized by the introduction of DeepSeek V3 and R1 models. These advanced language models represent significant milestones in natural language processing and reasoning capabilities. This article provides a detailed comparison of DeepSeek V3 and DeepSeek R1, exploring their features, performance, and practical applications.

Basic Introduction of Model

To begin our comparison, we first understand the fundamental characteristics of each model.

DeepSeek V3

• Release Date: December 27, 2024
• Model Scale:
  • deepseek/deepseek_v3
• Key Features:
  • Model Size: 671B parameters (37B active/token)
  • Tokenizer: SentencePiece-based multilingual tokenizer
  • Supported Languages: Focused on Chinese, English, and Japanese
  • Multimodal: Text-only
  • Context Window: 128K tokens
  • Storage Formats: FP8/BF16 inference
  • Architecture: Mixture of Experts (MoE) + Multi-Head Latent Attention
  • Training Method: Pre-training → Supervised Fine-Tuning (SFT) → Reinforcement Learning (RL)
  • Training Data: 14.8T tokens for pre-training

DeepSeek R1

• Release Date: January 21, 2025
• Model Scale:
  • deepseek/deepseek-r1
  • deepseek/deepseek-r1-distill-qwen-14b
  • deepseek/deepseek-r1-distill-qwen-32b
• Key Features:
  • Model Size: 671B parameters (37B active/token)
  • Tokenizer: Enhanced tokenizer with self-reflection tags
  • Supported Languages: Multilingual with cultural adaptation
  • Multimodal: Text-only
  • Context Window: 128K tokens
  • Storage Formats: Q8/Q5 quantization support
  • Architecture: Mixture of Experts (MoE) + RL-enhanced training pipeline
  • Training Method: Built on V3 base with RL pipeline (SFT → RL → SFT → RL)
  • Training Data: V3 base + RL optimization data

Model Comparison

Similarities:

• Both have the same model size (671B parameters, 37B active parameters per token).
• Both use the Mixture-of-Experts (MoE) architecture.
• Both are multilingual models excelling in English and Chinese.

Key Differences:

• Training Methods: V3 uses a traditional pipeline of pre-training, supervised fine-tuning (SFT), and reinforcement learning (RL). In contrast, R1 focuses on an RL-centric approach, incorporating cold-start fine-tuning and reward mechanisms to enhance reasoning capabilities.

Speed Comparison

If you want to test it yourself, you can start a free trial on the Novita AI website.

Speed Comparison

Cost Comparison

DeepSeek R1 surpasses DeepSeek V3 in output speed, but it has a longer total response time. The input and output prices of DeepSeek R1 are significantly higher than those of DeepSeek V3.

Benchmark Comparison

Now that we’ve established the basic characteristics of each model, let’s delve into their performance across various benchmarks. This comparison will help illustrate their strengths in different areas.

These results suggest that DeepSeek-R1 is better optimized for reasoning-intensive and domain-specific tasks (e.g., Codeforces and MATH-500), while DeepSeek-V3 delivers strong, though slightly lower, performance across these benchmarks.

If you want to see more comparisons, you can check out these articles:

• Deepseek v3 vs Llama 3.3 70b: Language Tasks vs Code & Math
• Llama 3.2 3B vs DeepSeek V3: Comparing Efficiency and Performance.

Applications and Use Cases

DeepSeek V3

• Designed for a broad range of tasks, including natural language understanding, coding, and basic problem-solving.
• Applicable across industries such as education, content creation, and business automation.
• Excels in domains like text generation, code completion, and mathematical reasoning.
• A versatile, general-purpose model suitable for various applications.

DeepSeek R1

• Tailored for tasks requiring advanced reasoning, logical inference, and mathematical problem-solving.
• Ideal for tackling complex, industry-specific challenges in fields like healthcare, finance, and legal services.
• Particularly effective for tasks demanding extended Chain-of-Thought (CoT) reasoning, such as diagnosing intricate problems, analyzing multi-step scenarios, and synthesizing insights from large datasets.

Accessibility and Deployment through Novita AI

Novita AI is an AI cloud platform that offers developers an easy way to deploy AI models using our simple API, while also providing the affordable and reliable GPU cloud for building and scaling.

Step 1: Log In and Access the Model Library

Log in to your account and click on the Model Library button.

Step 2: Choose Your Model

Browse through the available options and select the model that suits your needs.

Step 3: Start Your Free Trial

Begin your free trial to explore the capabilities of the selected model.

Step 4: Get Your API Key

To authenticate with the API, we will provide you with a new API key. Entering the “Settings“ page, you can copy the API key as indicated in the image.

Step 5: Install the API

Install API using the package manager specific to your programming language.

After installation, import the necessary libraries into your development environment. Initialize the API with your API key to start interacting with Novita AI LLM. This is an example of using chat completions API for python users.

from openai import OpenAI
  
client = OpenAI(
    base_url="https://api.novita.ai/v3/openai",
    api_key="<YOUR Novita AI API Key>",
)

model = "deepseek/deepseek_v3"
stream = True # or False
max_tokens = 2048
system_content = """Be a helpful assistant"""
temperature = 1
top_p = 1
min_p = 0
top_k = 50
presence_penalty = 0
frequency_penalty = 0
repetition_penalty = 1
response_format = { "type": "text" }

chat_completion_res = client.chat.completions.create(
    model=model,
    messages=[
        {
            "role": "system",
            "content": system_content,
        },
        {
            "role": "user",
            "content": "Hi there!",
        }
    ],
    stream=stream,
    max_tokens=max_tokens,
    temperature=temperature,
    top_p=top_p,
    presence_penalty=presence_penalty,
    frequency_penalty=frequency_penalty,
    response_format=response_format,
    extra_body={
      "top_k": top_k,
      "repetition_penalty": repetition_penalty,
      "min_p": min_p
    }
  )

if stream:
    for chunk in chat_completion_res:
        print(chunk.choices[0].delta.content or "", end="")
else:
    print(chat_completion_res.choices[0].message.content)
  
  

Upon registration, Novita AI provides a $0.5 credit to get you started!

If the free credits is used up, you can pay to continue using it.

DeepSeek V3 and DeepSeek R1 are powerful LLMs with distinct strengths. DeepSeek V3 is a versatile, general-purpose model known for its efficiency and strong performance across various tasks. DeepSeek R1, on the other hand, is a specialized model optimized for advanced reasoning. Choosing between them depends on the specific requirements of the application. Both models are significant advancements in the field, challenging existing models with their performance, efficiency, and open-source accessibility.

Frequently Asked Questions

Novita AI is the All-in-one cloud platform that empowers your AI ambitions. Integrated APIs, serverless, GPU Instance — the cost-effective tools you need. Eliminate infrastructure, start free, and make your AI vision a reality.

Recommend Reading

• DeepSeek V3: Advancing Open-Source Code Models, Now Available on Novita AI
• Deepseek v3 vs Llama 3.3 70b: Language Tasks vs Code & Math
• Llama 3.2 3B vs DeepSeek V3: Comparing Efficiency and Performance.