DeepSeek

DeepSeek Models: Comprehensive Guide & Usage Tutorial (2025)

This document provides a complete overview of the DeepSeek model matrix as of early 2025. It highlights the latest R1 Reasoning Series, V3 Flagship Series, and the Janus Multimodal Series, accompanied by detailed guides for API integration and local deployment.

---

1. Core DeepSeek Model Matrix

1. Flagship General & Reasoning Series (Current Mainstays)

This series represents DeepSeek's most powerful performance, setting the standard for open-source models globally.

• DeepSeek-R1 Series (Released Jan 2025)

  • Positioning: Specialized in Reinforced Reasoning, featuring "Chain of Thought (CoT)" capabilities similar to OpenAI's o1. It excels in complex logic, mathematical proofs, and deep thinking.
  • DeepSeek-R1: The full-scale model trained on the V3 architecture (671B parameters, 37B active). Its performance rivals GPT-4o and o1-mini, making it the strongest open-source reasoning model available.
  • DeepSeek-R1-Zero: The first-generation pure Reinforcement Learning (RL) model without Supervised Fine-Tuning (SFT). It demonstrates self-evolution capabilities, though its output style is rawer (ideal for research).
  • DeepSeek-R1-Distill (Distilled Versions): To adapt to consumer-grade hardware, DeepSeek "distilled" R1's reasoning capabilities into smaller models based on Qwen2.5 and Llama3:
    • 1.5B / 7B / 8B: Lightweight, suitable for laptops or standard desktops.
    • 14B / 32B: Mid-range. The 32B version is the "Sweet Spot" for value, performing exceptionally well on 24GB VRAM GPUs.
    • 70B: Heavyweight, offering performance close to the full-scale version.

• DeepSeek-V3 (Released Dec 2024)

  • Positioning: The third-generation flagship general-purpose model and the foundation for R1.
  • Features: Utilizes MoE (Mixture of Experts) architecture, supporting a 128K context window. Thanks to FP8 mixed-precision training, generation speed reaches up to 60 TPS (3x faster than V2).
  • Performance: Surpasses other open-source models in knowledge Q&A, long-text processing, and code generation. With extremely low API pricing, it is the top choice for general conversation.

2. Multimodal & Vision Series (Latest Breakthroughs)

• Janus-Pro Series (Released Jan 2025)

  • Positioning: A next-gen unified multimodal model capable of both "Seeing" (Understanding) and "Drawing" (Generation).
  • Features: Decouples visual encoding, ensuring that understanding and generation tasks do not interfere with each other.
  • Janus-Pro-7B: Built on DeepSeek-LLM-7B and SigLIP. It performs excellently in multimodal understanding benchmarks while supporting high-quality text-to-image generation.
  • Janus-Pro-1B: A lightweight version suitable for edge devices.

• DeepSeek-VL2

  • Positioning: A Vision-Language model focused on deep fusion of image and text understanding, ideal for OCR, chart analysis, and visual Q&A.

3. Coding & Professional Series

• DeepSeek-Coder V2

  • Positioning: The benchmark for code intelligence.
  • Features: Supports 338 programming languages. It excels in code generation, completion, bug fixing, and mathematical programming. While V3 is powerful generally, Coder V2 remains highly relevant for specific coding workflows.

• DeepSeekMath & DeepSeek-Prover

  • DeepSeekMath: An early model specialized in mathematics.
  • DeepSeek-Prover V1.5: Combines Reinforcement Learning with Monte Carlo Tree Search (MCTS), dedicated to formal theorem proving.

4. Historical Versions (Archived)

• DeepSeek-V2.5: Released Sept 2024, merging V2-Chat and Coder-V2. A transitional version.
• DeepSeek-V2: Introduced the DeepSeekMoE architecture, optimizing training costs.
• DeepSeek LLM (7B/67B): The early foundational large model series.

---

2. Usage Tutorial: How to Use DeepSeek

Method 1: Official Web / App (Easiest)

Suitable for general users; no technical background required.

• Web: Visit chat.deepseek.com
  • Register to use the V3 model for free.
  • Click the "DeepThink (R1)" button below the input box to switch to the R1 reasoning model.
• App: Search for "DeepSeek" in your mobile app store.

Method 2: API Integration (Recommended for Developers)

The DeepSeek API is fully compatible with the OpenAI format, allowing for a drop-in replacement.

Python Example:

# Prerequisite: pip install openai
from openai import OpenAI

client = OpenAI(
    api_key="YOUR_API_KEY",  # Apply at platform.deepseek.com
    base_url="https://api.deepseek.com"
)

response = client.chat.completions.create(
    # Model Selection:
    # "deepseek-chat"     -> DeepSeek-V3 (General Chat)
    # "deepseek-reasoner" -> DeepSeek-R1 (Deep Reasoning)
    model="deepseek-chat",
    messages=[
        {"role": "system", "content": "You are a professional assistant"},
        {"role": "user", "content": "Write a Snake game in Python"},
    ],
    stream=False
)

print(response.choices[0].message.content)

Method 3: Local Deployment (Privacy & Offline)

Using Ollama, you can run the distilled versions of DeepSeek-R1 locally on your computer. This requires no internet connection and ensures complete data privacy.

Steps:

1. Download Tool: Visit Ollama.com to download and install.
2. Select Model: Choose the version that fits your GPU VRAM:
   • 4GB - 8GB VRAM: Recommend 7B or 8B versions.
   • 12GB - 24GB VRAM: Recommend 14B or 32B versions.
3. Run Command: Open your terminal (CMD/Terminal) and enter the corresponding command:

# Run 1.5B Version (Ultra-low resource)
ollama run deepseek-r1:1.5b

# Run 8B Version (Mainstream recommendation, fast)
ollama run deepseek-r1:8b

# Run 32B Version (High performance, strong logic)
ollama run deepseek-r1:32b

---

3. Visual Interface (Optional)

For a user experience similar to ChatGPT but running locally, we recommend Chatbox AI.

Configuration:

• AI Model Provider: Ollama
• API Host: http://localhost:11434

---

📌 Summary: Which Model Should You Choose?

Select the model that best fits your specific use case: