Just yesterday, Deepseek quietly pushed out the software release of its latest model, the Deepseek-V3-0324. Based on early tests, V3 not only outweighs Openai’s GPT-4O and Sonnet 3.5 in coding and mathematics, but also retains its own against Claude 3.5 and 3.7. Mac Studio records 20 tokens per second. It’s not bad for something you can run locally.
There are no announcements. There is no hype. Just drops on your hugging face. But it didn’t remain quiet for a long time. For hours, users began to let it go through its pace. One outstanding feature? Ability to generate fully functional interactive websites from simple prompts.
Deepseek’s silent power play
Since January, Deepseek has been slowly rocking the AI scene after beating the likes of Meta, Openai and more on major benchmarks. With this new V3 update, the company is chasing AI-assisted coding. This is an area that is a playground for tools such as Cursor, Lovable, Windsurf.
And that’s not just a story. One user from x, @el.cine, in Paris, urged deepseek to:
“You are a website developer. I code the latest little digital marketing landing page.”
result? A sophisticated work site – built without touching a line of code. Just a prompt.
Openai’s Thisis Nightmare
China has just dropped the DeepSeek V3-0324, and is a 100% open source AI model with just 700GB
– Better than Sonnet 3.5
– Competing with Claude 3.7 in Coding & Math
– Run 20 tokens/s in Mac StudioYou can generate Interactive website (comment code) pic.twitter.com/6bxocgpwdw
– el.cine (@ehuanglu) March 25, 2025
I ran the same test. DeepSeek generated the complete source code within 5 minutes. HTML, CSS, JavaScript – Clean, production compatible. We rolled out it with a zero adjustment and it looked like what you would pay the developer hundreds of dollars to spin up. I gave DeepSeek the same prompt. This is what was built. Try it yourself with Deepseek.
From prompts to work websites – no code required
Looking at the code is one thing. It’s a different matter to see tools that interpret high-level instructions and turn them into usable ones. That’s exactly what DeepSeek-V3 is doing. This experience felt like it wasn’t programming, but rather giving instructions to someone who had already got what you were trying to build.
This is what excites developers. X’s post is full of screenshots and praise: fast generation, clean UI, amazingly good code. And it’s free. The combination puts pressure on premium platforms that charge with less charge.
Vibe Coding is here and Deepseek owns it
The term coding for atmosphere may not be in textbooks yet, but it’s catching up. It’s the idea that you can sketch concepts in natural language, and AI handles the details. There are no syntax errors. There are no boiler plates. Just make your code work with a high level of intent.
Deepseek’s models, especially V3 and R1, are built for this type of interaction. They interpret their intentions rather than simply spitting out the code. Whether it’s a marketing site or a basic app, Deepseek responds like you know what it means. It is conversational and accurate in a way that feels natural.
But what really separates Deepseek from the rest is the price. Openai charges $200 a month to access the largest models, while DeepSeek offers comparable output for free. Training cost? Only $5.6 million. Comparing it to its hundreds of millions of competitors, we start to see why this is important.
Openai or a friend’s trouble?
This isn’t just about creating a cool website. Deepseek’s Rise is threatening the biggest players in this space. Companies like Openai, Google and Anthropic are building moats on premium models and expensive infrastructure. Deepseek is destroying that pattern with cheaper, faster, open source alternatives.
Architecture is important here. Deepseek uses mixture use and multi-token prediction to achieve speed and efficiency without the large hardware requirements. In addition, you can generate complete artifacts (Webpages, UIS, working components) in real time.
Back up your benchmarks. The R1 surpasses Openai’s O1 in tests such as the Math-500 and SWE bench. And don’t forget about the market reaction in January: Nvidia’s stock lost $600 billion worth of $600 billion in a day when Deepseek’s app hit No. 1 on the App Store and investors feared reforms in computational demand.
Openai criticized Deepseek for distilling the model, but the bigger story is the response from users and how quickly they are appearing.
Cursors, adorable, windsurf – who is at risk?
Cursors, bolts, lovable, windsurf – all of them are now facing real pressure. These platforms gained traction by facilitating coding with AI agents and smooth interfaces. But Deepseek offers many of the same value at no charge.
For example, Cursor charges $20 a month to access premium features such as composers and agent tools backed by Claude and Openai APIs. However, users are adding DeepSeek to Cursor via browser extensions like Cline. Skiing fully with subscription costs. The power of the model is becoming a product.
The bolts and adorable ones are also in a pinch. They shine with quick prototyping and quick front end, but DeepSeek does the same with two prompts. As one post on x said:
“Bolt’s $105 million Raise feels unstable if Deepseek can build the same app for free at two prompts.”
Even Windsurf, known for its sophisticated cascades, is not immune. DeepSeek-R1 is already integrated through Codeium. It says a lot: If the raw model is strong enough, what is the value of the platform around it?
Windsurf’s $10/month subscription will start to soar next to Deepseek’s open ecosystem, especially if these models can be run locally through tools such as Ollama or VS code integration. Deepseek is not only competing, it is evacuating.
This is not just a moment, it’s momentum
The gaps expand faster as the Deepseek community continues to grow and developers continue to build wrappers, plugins and tutorials. Cursors, bolts, and more should lean violently towards a great user experience and pivot for enterprise customers. Otherwise they risk replacing it with the very thing they popularized.
Where Deepseek goes from here
The latest V3-0324 release is just the beginning. Developers are already pushing the limits from creating full stack apps to generating internal tools in record time. For startups and indie developers, this could mean cost reductions and speed. For businesses, that may mean rethinking their dependence on closed platforms.
And for the incumbent? The pressure is real. They either need to innovate quickly or risk seeing leads escape.
When one user summarises it in X:
“Deepseek has been released to people who are almost unexpected. A reasoning model equivalent to Openai’s O1… It feels like things are trying to speed up in the AI world.”
Two prompts, one wake-up call
What started as a simple, quick test has turned into something more. This looks at what AI development will look like soon. Deepseek hasn’t caught up. You are turning the script over. Also, if you’re building software, or building tools that help others build it, be aware of what happens next.
Web pages generated with two DeepSeek prompts (without manual coding).
