No press conference, no overwhelming publicity, the DeepSeek V3-0324 version was quietly launched on March 24.

https://huggingface.co/deepseek-ai/DeepSeek-V3-0324

Although this update does not belong to the new generation of R2, its improvement is far beyond expectations, especially in terms of programming capabilities, front-end UI design and context understanding, it is already comparable to Claude 3.7 Sonnet.

Another important upgrade is that DeepSeek continues to adhere to the spirit of open source and directly adopts the MIT open source protocol, allowing developers to freely modify and commercially use it, completely breaking the monopoly of closed source AI on top code models.

What does all this mean?

What has DeepSeek V3-0324 evolved?

DeepSeek V3 itself is a highly cost-effective open source model recognized by the industry. Although the 0324 version is only a "small version update", its improvement is not small. That's right, DeepSeek is so low-key.

The parameter scale has been further improved: the model parameters have increased from 671 billion to 685 billion. Although the increase is not large, the optimized computing efficiency has brought smoother reasoning capabilities.

Coding capabilities have improved by leaps and bounds: The new DeepSeek V3's capabilities in front-end development, website construction, and UI design have been significantly improved. Everyone basically believes that it can be compared with Claude 3.7 Sonnet.

Mathematical reasoning is stronger: the ability to improve multiple rounds of dialogue and logical reasoning makes its performance in answering mathematical problems more stable, and some problem-solving abilities are close to the reasoning model.

Open source license is more free: the adoption of MIT licenses allows developers to modify, optimize and commercially use it at will, greatly reducing the threshold for AI applications.

Practical performance: Is the new version of V3 really so capable?

In the test of professional users, the code capability of DeepSeek V3-0324 has reached the world's strongest open source AI level, and has even surpassed Claude 3.7 in some scenarios.

Some netizens said that in the Aider benchmark test, DeepSeek V3-0324 has significantly improved, close to Claude Sonnet 3.7.

As some netizens released test results for front-end design, which can be seen that the design capabilities of the new version of V3 even exceed R1.

The following example is more obvious, and we directly compare several top models. This netizen asked them to generate codes for animated weather cards.

The results are very intuitive. The effects of the new V3 are far superior to the original V3 and R1, and are directly comparable to Claude 3.7.

A AI developer sighed: "Since June 2024, no LLM on the market has been close to Claude 3.7 in front-end code generation, but DeepSeek V3-0324 has finally done it."

Open source killer: MIT license, AI enters the era of freedom?

The biggest "hidden easter egg" in this update is that DeepSeek completely unlocks the MIT license, making V3-0324 a completely open and commercial top AI model.

Why is MIT license so important?

Commercial use is not hindered: both enterprises and individual developers can use it freely without paying expensive licensing fees.

Free transformation of the model: it can be modified, fine-tuned at will, and even used in its own products or projects without official restrictions.

Promote the development of the AI ​​ecosystem: Compared with OpenAI's closed-source business models such as GPT-4.5 and Claude 3.7, DeepSeek allows more entrepreneurial teams and individual developers to enjoy top AI capabilities equally, truly realizing "technical equality".

A foreign netizen said: "DeepSeek is doing what Meta promised but did not, allowing AI to really move towards open source and freedom. ”

But some people are happy and some are worried. If open source AI like DeepSeek continues to soar, what else can closed source AI rely on to attract users?

The AI ​​competition is getting hot, and the business model is "shivering"

The release of DeepSeek V3-0324 not only enhances the status of open source AI in the fields of programming and mathematics, but also brings unprecedented pressure to closed source AI companies such as OpenAI and Anthropic.

Previously, closed-source AI could still attract high-end users with its "stronger capabilities", but now, the free and open-source DeepSeek V3 has approached Claude 3.7, and is even better in some areas.

A foreign user bluntly said: "Seeing the advancement of these open-source models, I really hope that commercial AI companies can reduce prices. ”

DeepSeek's upgrade has also made people begin to re-examine the future of the AI ​​industry. When high-performance AI gradually moves towards open source, does the AI ​​business model need to be completely changed?

V3 evolves, will R2 be far away?

Then, the next question is - will R2 be far away?

Many DeepSeek users said that If V3-0324 is already so powerful, then R2 may really be about to overturn the entire AI ecosystem.

In the next few months, the AI ​​field may usher in greater changes, and DeepSeek has already stood at the center of the storm.