deepseek-ai/DeepSeek-V3-0324 · Review of DeepSeek V3 0324

Among freely available LLMs, this is one of the best models, but you need to know how to properly configure it and where to find it.

First, locate it:
https://openrouter.ai/deepseek/deepseek-chat-v3-0324:free

Sample results:
https://disk.yandex.ru/d/VZkFsMwh8bwNjQ
Check "Variant 4".

All results are significantly better:
https://disk.yandex.ru/d/iP_f37VTFKm_rA

Optimal parameters for perfect output:

Temperature: 0.6  
Top P: 0.95  
Top K: 30  
Min P: 0.03

Comment

Prompt:
"Write complete Python 3.11 code for Windows using OpenCV to implement the following algorithm:

The program should load a video file specified via a command-line argument.
Compute the horizontal displacement of a moving object between frames using the ORB algorithm. Implement a filtering mechanism to exclude points unrelated to the moving object. Prioritize points with the fastest motion, where their velocities differ by ≤1 pixel/frame, and ensure at least 3 such points meet this criterion.
Generate an image composed of vertical line segments from each frame. The width of each segment corresponds to the displacement magnitude from step 2. Apply horizontal flipping if the displacement is positive. Each segment starts at the frame’s center and extends rightward by the displacement length.
Save the final composite image."

All newer free models underperform. The more powerful Google Gemini 2.5 PRO (free) became unavailable after April 17, 2025, despite being labeled "free."

April 20, 2025 Update
After testing DeepSeek R1 Zero (see: https://nikitayev.livejournal.com/157013.html), I experimented further with DeepSeek V3 0324 and found more stable parameters:

Temperature: 0.0  
Top P: 0.0 (or 1.0)  
Top K: 100  
Min P: 0.00

Stability and quality improve only when using a system prompt that forces the model to "think" like R1.

Simplest effective system prompt:

# You are a world-class AI system capable of complex reasoning and reflection.  

## You respond to all questions in the following way:  

    <think>  
        - Understand the problem and devise a solution plan.  
        - For simple problems: Use straightforward Chain of Thoughts.  
        - For complex problems:  
            1. Create a step-by-step plan (do not execute yet).  
            2. Break down the problem, use best practices for math/programming, and validate solutions.  
            3. Use <reflection> tags to verify critical steps.  
    </think>  

    <output>  
        - Provide the full answer based on your reasoning.  
    </output>

This prompt is available here:
https://disk.yandex.ru/d/iP_f37VTFKm_rA
Folder: «Промпты\Системные промпты\Универсальные промпты», file: «Промпт 23 reflection ENG.txt»

Why DeepSeek V3 remains useful:

V3 can resume generation by typing "continue," whereas R1 restarts from scratch.
For unrestricted R1 Zero usage, consider running it locally on a server with 1TB RAM.

Translation Note: URLs and folder/file names in Cyrillic are preserved as they are part of external links.