I Tested DeepSeek’s R1 and V3 Coding Skills – and we’re not All Doomed (Yet).

DeepSeek exploded into the world’s awareness this past weekend. It stands apart for three effective reasons:

1. It’s an AI chatbot from China, instead of the US

2. It’s open source.

3. It uses greatly less facilities than the huge AI tools we’ve been looking at.

Also: Apple scientists reveal the behind DeepSeek AI

Given the US government’s issues over TikTok and possible Chinese government involvement in that code, a new AI emerging from China is bound to create attention. ZDNET’s Radhika Rajkumar did a deep dive into those issues in her post Why China’s DeepSeek might burst our AI bubble.

In this post, we’re avoiding politics. Instead, I’m putting both DeepSeek V3 and DeekSeek R1 through the same set of AI coding tests I’ve thrown at 10 other large language models. According to DeepSeek itself:

Choose V3 for tasks requiring depth and precision (e.g., resolving innovative math issues, producing intricate code).

Choose R1 for latency-sensitive, high-volume applications (e.g., customer support automation, standard text processing).

You can select between R1 and V3 by clicking the little button in the chat user interface. If the button is blue, you’re using R1.

The brief response is this: remarkable, however clearly not ideal. Let’s dig in.

Test 1: Writing a WordPress plugin

This test was in fact my very first test of ChatGPT’s shows prowess, method back in the day. My wife required a plugin for WordPress that would help her run an involvement device for her online group.

Also: The very best AI for coding in 2025 (and what not to use)

Her requirements were relatively simple. It needed to take in a list of names, one name per line. It then had to sort the names, and if there were replicate names, different them so they weren’t listed side-by-side.

I didn’t truly have time to code it for her, so I chose to offer the AI the obstacle on a whim. To my huge surprise, it worked.

Ever since, it’s been my very first test for AIs when assessing their programming skills. It needs the AI to know how to set up code for the WordPress structure and follow prompts clearly adequate to create both the user interface and program reasoning.

Only about half of the AIs I’ve checked can totally pass this test. Now, however, we can include another to the winner’s circle.

DeepSeek V3 developed both the user interface and program logic exactly as specified. When It Comes To DeepSeek R1, well that’s an interesting case. The „reasoning” aspect of R1 triggered the AI to spit out 4502 words of analysis before sharing the code.

The UI looked different, with much larger input areas. However, both the UI and logic worked, so R1 also passes this test.

Up until now, DeepSeek V3 and R1 both passed one of four tests.

Test 2: Rewriting a string function

A user grumbled that he was unable to go into dollars and cents into a donation entry field. As composed, my code just permitted dollars. So, the test includes offering the AI the regular that I wrote and asking it to reword it to allow for both dollars and cents

Also: My preferred ChatGPT function just got way more powerful

Usually, this results in the AI creating some regular expression validation code. DeepSeek did produce code that works, although there is room for improvement. The code that DeepSeek V2 composed was needlessly long and repetitious while the reasoning before creating the code in R1 was likewise long.

My most significant concern is that both models of the DeepSeek recognition ensures recognition approximately 2 decimal locations, but if a large number is gone into (like 0.30000000000000004), using parseFloat doesn’t have specific rounding knowledge. The R1 model likewise utilized JavaScript’s Number conversion without checking for edge case inputs. If bad data comes back from an earlier part of the routine expression or a non-string makes it into that conversion, the code would crash.

It’s odd, because R1 did provide a very great list of tests to confirm against:

So here, we have a split decision. I’m providing the indicate DeepSeek V3 due to the fact that neither of these problems its code produced would cause the program to break when run by a user and would produce the anticipated results. On the other hand, I need to give a fail to R1 because if something that’s not a string somehow enters into the Number function, a crash will ensue.

And that offers DeepSeek V3 two triumphes of 4, but DeepSeek R1 just one triumph of four up until now.

Test 3: Finding an annoying bug

This is a test produced when I had a very irritating bug that I had problem locating. Once once again, I decided to see if ChatGPT might handle it, which it did.

The obstacle is that the response isn’t apparent. Actually, the obstacle is that there is an apparent response, based upon the error message. But the apparent response is the wrong answer. This not only caught me, but it regularly captures a few of the AIs.

Also: Are ChatGPT Plus or Pro worth it? Here’s how they compare to the totally free version

Solving this bug needs comprehending how specific API calls within WordPress work, having the ability to see beyond the mistake message to the code itself, and then understanding where to find the bug.

Both DeepSeek V3 and R1 passed this one with nearly identical answers, bringing us to three out of four wins for V3 and 2 out of 4 wins for R1. That currently puts DeepSeek ahead of Gemini, Copilot, Claude, and Meta.

Will DeepSeek score a home run for V3? Let’s learn.

Test 4: Writing a script

And another one bites the dust. This is a difficult test because it requires the AI to comprehend the interplay between three environments: AppleScript, the Chrome things model, and a Mac scripting tool called Keyboard Maestro.

I would have called this an unjust test due to the fact that Keyboard Maestro is not a mainstream programming tool. But ChatGPT dealt with the test quickly, understanding exactly what part of the problem is managed by each tool.

Also: How ChatGPT scanned 170k lines of code in seconds, conserving me hours of work

Unfortunately, neither DeepSeek V3 or R1 had this level of knowledge. Neither design understood that it needed to divide the job in between directions to Keyboard Maestro and Chrome. It also had relatively weak knowledge of AppleScript, writing customized regimens for AppleScript that are native to the language.

Weirdly, the R1 design stopped working also since it made a bunch of inaccurate assumptions. It assumed that a front window constantly exists, which is definitely not the case. It also made the assumption that the presently front running program would always be Chrome, instead of clearly checking to see if Chrome was running.

This leaves DeepSeek V3 with three right tests and one fail and DeepSeek R1 with 2 appropriate tests and 2 fails.

Final thoughts

I discovered that DeepSeek’s insistence on utilizing a public cloud email address like gmail.com (instead of my typical e-mail address with my corporate domain) was bothersome. It likewise had a number of responsiveness stops working that made doing these tests take longer than I would have liked.

Also: How to use ChatGPT to write code: What it does well and what it does not

I wasn’t sure I ‘d have the ability to write this short article because, for most of the day, I got this error when trying to register:

DeepSeek’s online services have recently faced large-scale destructive attacks. To make sure continued service, registration is momentarily restricted to +86 contact number. Existing users can visit as typical. Thanks for your understanding and assistance.

Then, I got in and was able to run the tests.

DeepSeek appears to be excessively loquacious in terms of the code it generates. The AppleScript code in Test 4 was both incorrect and excessively long. The regular expression code in Test 2 was correct in V3, however it might have been composed in a way that made it much more maintainable. It failed in R1.

Also: If ChatGPT produces AI-generated code for your app, who does it really come from?

I’m absolutely satisfied that DeepSeek V3 beat out Gemini, Copilot, and Meta. But it appears to be at the old GPT-3.5 level, which means there’s definitely space for improvement. I was disappointed with the results for the R1 model. Given the option, I ‘d still pick ChatGPT as my shows code helper.

That stated, for a brand-new tool working on much lower facilities than the other tools, this could be an AI to enjoy.

What do you believe? Have you tried DeepSeek? Are you using any AIs for shows support? Let us know in the remarks below.

You can follow my everyday task updates on social networks. Make certain to register for my weekly upgrade newsletter, and follow me on Twitter/X at @DavidGewirtz, on Facebook at Facebook.com/ DavidGewirtz, on Instagram at Instagram.com/ DavidGewirtz, on Bluesky at @DavidGewirtz. com, and on YouTube at YouTube.com/ DavidGewirtzTV.