Beyond Imagination! DeepSeek V3 Minor Update: Shockingly Powerful in Real-World Tests, Outperforms R1, and Rivals Claude 3.7

As Dr. Kai-Fu Lee once said, “Sam Altman might lose sleep again!” Here’s why: DeepSeek has quietly rolled out a new version update—DeepSeek V3 0324—without any fanfare, benchmark announcements, or marketing hype. The model, approximately 700GB in size, was directly uploaded to Hugging Face under an MIT license, meaning it can be freely deployed or tested right now on DeepSeek’s official website (just avoid enabling “thinking mode”).

At first glance, this seemed like just another routine update, so I didn’t pay much attention. But after running some tests, I was blown away—this version is absurdly powerful! My hands-on experience suggests its coding capabilities, especially in frontend development, are on par with Claude 3.7 Sonnet. Considering DeepSeek’s competitive API pricing, this is nothing short of astonishing!!! Below are my real-world tests—all executed in a single attempt without any modifications.

Test 1: Cyberpunk Snake Game

This is one of my go-to prompts for evaluating the latest versions of various LLMs.

Prompt:
*”Help me create a Cyberpunk-themed Snake game that runs in a single HTML file.”*

Test Process: [Duration: 00:12]

Results:
The implementation had no logical errors and effectively captured the Cyberpunk aesthetic. Surprisingly, it outperformed both R1 and Claude 3.7 in execution quality (I’ve documented their performances in previous articles). While a Snake game may seem simple, adding “Cyberpunk” as a constraint often trips up even newer models, introducing bugs—so I won’t dwell on comparisons here.

[Duration: 00:19]

Test 2: Animated Weather Card

This test was based on a user-submitted prompt to see if the model could replicate it accurately.

Prompt:
*”Create a single HTML file containing CSS and JavaScript to generate an animated weather card. The card should visually represent the following weather conditions with distinct animations:

• Wind (e.g., moving clouds, swaying trees, or wind lines)
• Rain (e.g., falling raindrops, puddles forming)
• Sun (e.g., shining rays, bright background)
• Snow (e.g., falling snowflakes, snow accumulating)
  Show all weather cards side by side on a dark background. Include all HTML, CSS, and JavaScript in one file, with a way to toggle between weather conditions (e.g., buttons or a function) to demo each animation.”*

Test Process: [Duration: 00:13]

Results:
Flawless execution! The design was polished, packed with details like dynamic snow accumulation and highlighted weather-toggle buttons. The only minor shortcoming was that the cards weren’t perfectly aligned side-by-side, but overall, its instruction-following ability was exceptional—even surpassing Claude 3.7.

[Duration: 00:15]

For comparison, here’s Claude 3.7 Sonnet’s attempt—notice the missing details and lack of toggle buttons.

[Duration: 00:04]

I also tested R1 with the same prompt, and its output was clearly inferior to V3’s new version.

[Duration: 00:16]

Test 3: Solar System Simulation

This test evaluates the model’s ability to infer complex systems from minimal instructions.

Prompt:
*”Simulate a solar system with planetary motion in a single HTML file.”*

Test Process: [Duration: 00:09]

Results:
Absolutely mind-blowing! Despite the vague prompt (no specifics about orbits, planets, or cosmic environment), the model simulated planetary motion with three adjustable modes: Slow, Standard, and Accelerated. Saturn even had its iconic rings—an unexpected yet brilliant touch!

[Duration: 00:32]

R1’s attempt with the same prompt paled in comparison, lacking many of these refined details.

[Duration: 00:15]

Final Thoughts

My testing was quick and unsystematic, but early feedback from other users suggests DeepSeek V3 0324 also excels in mathematical tasks. You can try it yourself right now on DeepSeek’s official website.

Personal takeaway: A “minor” update with MAJOR leaps! This only heightens the anticipation for DeepSeek V4 and R2.