DeepSeek V3 vs GPT-4o: Which is better?

In a year marked by notable OpenAI releases like GPT-4o, Sora, and o1, the open source community has steadily moved forward with its own impressive models. Among them is DeepSeek V3, an open source model developed with a budget of just $5.5 million, significantly less than the training cost of GPT-4o (over $100 million). With 685 billion parameters and a mix of experts (MoE) architecture, DeepSeek V3 competes strongly in areas such as encoding and translation, while offering cost-effectiveness and open source flexibility. Let's explore how DeepSeek V3 compares to GPT-4o and what it brings to ai development.

Task 1: Create a flowchart

Immediate: Create a complete flowchart illustrating the reflection pattern in agent ai.
These are the steps involved in the reflection pattern:

Generate initial output
Self-assessment/criticism of the result
Identify errors, gaps or areas for improvement
Develop suggestions for improvement
Review/refine the result
Repeat steps 2 to 5 until you achieve a satisfactory result
Create a new flowchart now

DeepSeek V3 Output:

+---------------------+
| Generate Initial Output |
+---------------------+
|
v
+---------------------+
| Self-Review/Critique the Output |
+---------------------+
|
v
+---------------------+
| Identify Errors, Gaps, or Improvement Areas |
+---------------------+
|
v
+---------------------+
| Develop Improvement Suggestions |
+---------------------+
|
v
+---------------------+
| Revise/Refine the Output |
+---------------------+
|
v
+---------------------+
| Is the Result Satisfactory? |
+---------------------+
|
+--------+--------+
|                 |
v                 v
Yes                No
|                 |
v                 v
+---------------------+
| Final Output        |
+---------------------+
|
v
+---------------------+
| End                 |
+---------------------+

GPT 4o output:

Observation:

The flowchart provided by V3 is more effective than that of GPT-4o because it clearly describes the actions to be taken based on whether the result is satisfactory or not.

Verdict:

DeepSeek V3 | GPT 4o

Task 2: Zebra Puzzle

The first task of this GPT 4o vs DeepSeek V3 guide, I have a zebra puzzle of this website.

Immediate: Solve this zebra puzzle and give me a table of final results.

DeepSeek V3 Output:

Putting this answer on the website:

GPT 4o output:

Put this solution on the website:

Observation:

While both models assigned random names to elements where information was not available, V3 successfully solved the problem, while GPT-4o could not.

Verdict:

DeepSeek V3 | GPT 4o

Task 5: Physics Circuit Problem

Immediate: The figure shows part of a circuit. It consists of resistors combined in parallel and series configurations. Find the equivalent resistance.

DeepSeek V3 Output:

GPT 4o response:

Observation:

Comparing DeepSeek V3 and GPT-4 solutions for the given resistor network, GPT-4's calculation of 1.29 Ω is correct, while DeepSeek V3's result of 3.59 Ω is incorrect. GPT-4 correctly identified the circuit structure with three parallel branches: (R1+R2=3Ω), R3=3Ω and (R4+R5=9Ω), then accurately applied the parallel resistance formula (1/Rt = 1/ 3 + 1/3 + 1/9 = 7/9) to arrive at the final result. DeepSeek V3 made critical errors by incorrectly grouping resistors, misidentifying series and parallel combinations, leading to an inaccurate final calculation.

Verdict:

DeepSeek V3 | GPT 4th

Task 4: Article Summary

Immediate: Read the article at https://www.analyticsvidhya.com/blog/2024/07/building-agent-rag-systems-with-langgraph/ to understand the process of creating a vector database for Wikipedia data. Then, provide a concise summary of the key steps.

DeepSeek V3 Output:

GPT 4o output:

Observation:

DeepSeek V3's explanation is more complete and technically accurate, covering preprocessing, indexing, and LangGraph integration, along with recommendations for specific tools such as FAISS and Pinecone. The GPT-4 response, while clear and well-structured, omits critical technical elements and simplifies complex processes. DeepSeek V3's comprehensive coverage and technical depth make it most valuable for practical implementation, although GPT-4 excels at presenting information in an accessible format.

Verdict:

DeepSeek V3 | GPT 4o

Task 5: Find differences

Immediate: The image is divided into two parts that are almost identical. However, there are three elements present in the image on the left that are missing from the one on the right. Your task is to identify these missing elements.

DeepSeek V3 Output:

GPT 4o output:

Observation:

V3 was unable to parse the image directly and provided a generic response. GPT-4 identified one correct difference, but the remaining differences it suggested were incorrect.

Verdict:

DeepSeek V3 | GPT 4o

GPT 4o vs DeepSeek V3: final result

Also read:

Final note

DeepSeek V3 demonstrates that open source models can compete or even outperform commercial models like GPT-4o, while costing significantly less to train ($5.5 million vs. $100 million or more). Its strong performance and free accessibility make it a great choice for both developers and organizations looking for powerful ai capabilities without commercial restrictions.

I am very excited to use DeepSeek V3 and explore its full range of features. What about you? Share your thoughts in the comments below!