New DeepSeek R1 is Really, Really Good Coder

We all were waiting for R2 but the deep seq just dropped an upgrade to R1 and this thing is really good at coding. Here's an example of the previous generation and now you can generate this with exactly the same prompt which shows a huge improvement on coding. This thing is really good at one shot generation and does a really good job compared to the previous version. But it's also a kind of overthinker. So in this video we are going to look at how good this model is and why they probably did not release R2 yet. So the model weights are available in hugging face but if you want to test it out you can already check it out on deepseek website. If you're looking for us based hosting service openrouter is hosting it. Before testing the model let's look at the benchmarks that are available. So the scores on live code bench are already available and this thing is really impressive. So it's on par with O3 high. Very close here and some of my testing actually showed that it's a much better version compared to the original R1. This seemed to be built on top of the new V3 and it actually shows in the UIs that it creates. You can push it to create some very interesting looking UIs which we were not able to do with the original R1. Let's test this model out. Also I'm going to point out a couple of really interesting things that Deepseek has done with this upgraded R1. For testing the model we're going to be using Deepseek's own website. I'm going to test it on a number of different prompts and the results are really interesting to say the least. It also has the ability to use the web search tool, but the feature that I personally like a lot is that Deep SEQ is still showing raw chain of thoughts, which is pretty great because you can learn a lot from these raw chain of thoughts. Unfortunately, other frontier labs like Google with their Gemini OpenAI with their O series and now Claude has moved away from showing raw chain of thoughts to the user and they are showing summarized versions which are not really useful. I had a conversation with Sam about the importance of chain of thought or raw chain of thought tokens if you're interested. Make sure to watch this video. Now when it's generating these raw chain of thoughts, it seems to be a lot more organized compared to the previous R1. For example, you can see here that it's thinking about three different options and then based on that it's making A more educated decision, which option to adopt. Second, it thinks for a lot longer. I have seen it think for almost 10 minutes for certain test prompts. Okay, so here's a previous version of the same prompt. It took about 250 seconds. That's only the chain of thought. It also tends to implement the actual code in the chain of thoughts. This is something I have observed in the Gemini models, but not with the other thinking models. The implementation took, I think, about 30 more seconds. Okay, you can download the code or run it directly here. In here, it created the animation based on our requirements. Letters of falling down under the effect of gravity, and they also have realistic collisions. Let me actually download this file and run it. Okay, so it's keeping track of how many letters we can see on the screen, also the frame rate. And this seems to be doing really good. This was one of the prompts that the original O couldn't get. Okay, here's another prompt. So this is a variation of the viral prompt Ball falling within hexagon. But in this case, we want to have 20 bouncing balls spinning within heptagon. Now, the only model that can consistently get this right is Claude 3.7 Sonnet. I haven't tested it with Claude 4, but even Gemini 2.5 is not able to do this. And this new model is able to consistently generate really good results. The balls starts in the middle. They are colliding with each other. They are numbered correctly. I also added this couple of controls where you can spin it in the reverse direction, counterclockwise or clockwise. Also you can speed up, speed down, and it seems to actually obey the rules that we define. So really impressive. Here's another prompt that I used for testing both the coding capabilities as well as the creativity. So it's supposed to generate nine or ten different TV channels inspired from different genres of TV channels. We don't want it to use any HTML code. Everything is supposed to be within that sketch. Also the names to be creative, and then there needs to be animation aligned with the names. Okay, so here is what it came up with. That's channel number one. Channel two is news. So it seems like this is somebody who is talking behind a news desk. Cartoon Network. This actually looks pretty cute. Sports Zone. I think it could have done a much better job. I think Gemini 2.5 was the one that did the best in my experiments. This is news, music, cooking show, Sci Fi, Nature world, and then classic movie. And if you press on 1 or 0, we can go back. So it does a pretty decent job at this Prompt. Here's another one that basically tests in its internal knowledge and training data because in this case it doesn't have access to Internet. So basically we wanted to create a simple landing page of the first 25 legendary Pokemon and I wanted to see whether it has access to the web links and seems like it does. Here's the first version that it created. It fulfills all the requirements that we had in the prompt and also you can actually filter out these Pokemons, which is pretty neat. It also has the search capability and then I simply asked it to make it more futuristic, inspired by cyberpunk and you can see the results are really impressive. So here's the website. Now again it has all those different features. It adds these really neat animations. So just like the upgraded V3, this does, I think a really good job at creating Web apps and UIs in general. Now I want to test its ability to create interactive web apps. So I asked it to create a view of night sky with the top constellations and again it did, I think a really good job at it. So these are the constellations. We can rotate them, we can stop the rotation, you can hide the names. I think it has a lot of creativity. Since this is a reasoning model, I do want to test its reasoning capabilities for the that we're going to use a couple of prompts from the misguided attention repo. I have covered this in my previous videos. The first one is a modified version of the Trolley problem where five people on the track are dead. So it starts off by looking at the user presents the classic trolley problem scenario. This is well known thought experiments in ethics. But I need to approach it carefully since real people might be genuinely wrestling with the moral dilemma. Right. And then it says let's unpack what is being asked. Trolley heading towards five corpuses versus one living person. Right. Then it says that corpus detail is unusual. Most versions say five living people. So interestingly it's smart enough to figure out that we are talking about dead people. So it says the five are already dead. That fundamentally alters the situation. Pulling the lever won't save lives. So really smart. And you can actually see that it's a thinking process. It also mentioned Philipa Foot who seems to be the creator of this Trolley problem. I haven't seen this being shown by any LLMs. So this is something that I actually Learned from Deepsea Car 1 updated. Right. So it basically goes through this thought process and at the end it says the answer is clearly no. And here is the reasoning behind it. Right. So really smart so far. But then I tested on this. So this is basically a modified version of the famous farmer problem where the farmer has to take a wolf, a goat and a cabbage to the other side in the original statement. But in this case we only want the farmer to take the goat to the other side. That's it. We don't want him to take anything else to the other side. And this is where I think it's going back to its training data rather than actually doing logical deductions. So it started off by listing different item items that we need to do. Then it says let's call the starting site site A, side B, right? And here it is going through the different steps that it's going to take so that it can take everything to the other side rather than just focusing on the goat that we mentioned. And you can see it's reasoning through the different combinations of different steps, right? It is trying to figure out what are the few number of steps that it will be able to complete this whole transition. But it misses on the main thing, that we only want the goat to go to the other side rather than all of the items. And in the final answer you can actually see that happening. So it starts off by taking the goat across, then it thinks about how it's going to take the wolf and the cabbage and so on and so forth. So based on my quick testing, it's definitely an improvement or the original R1, especially when it comes to coding. Chain of thought seems to be also a lot more organized now, although it takes much longer compared to the original R1, but it tends to overthink a lot. So here is a quick example. Let's say if you ask it something like what is the capital of Pakistan? It says that seems straightforward. Let me start by recalling the capital Pakistan's capital is as. Right, so this is where it should stop and just generate the answer because the user is specifically asking about capital of Pakistan. But then it starts going to the history of how it moved from one city to another, right where it's located, and so on and so forth. So it's consuming a lot of thinking tokens, which I don't think you need for a problem like this. The thinking here can take a lot of time, so here it's almost five minutes. But the neat thing is you can disable the thinking tokens if you want. So there is a PR out which essentially converts this reasoning model into a non reasoning model. It's very similar to the feature that Gemini is shipping where you can disable the thinking token or enable specific thinking budget. This will make it extremely useful. So you can use exactly the same model both as a reasoning and non reasoning model. But keep in mind this is an independent pr so I'm not sure this is going to exactly work, but we are going to find out pretty soon now. So far we only have benchmarks for this live code bench because people are able to test it. Deepseek has not officially released any benchmarks. They only dropped the weights on hugging face and I feel like this probably was supposed to be R2. Maybe they haven't figured out a huge improvement compared to what they have, but we'll have to see because if you look at the original V3 and the updated V3 there was definitely an improvement between these two models. V3 was actually doing much better on coding compared to the original V3, so I feel like they just a dated it based on v3. R2 might be entirely different architecture and we probably are going to see it in the next month or so. Anyways, this was a quick video on this updated R1. It's definitely a really good model when it comes to coding. And let me know if you are interested in a comparison of this model compared to something like Gemini 2.5 Pro or the new Card 4 models. I can do a comparison anyways. I hope you found this video useful. Thanks for watching and as always, see you in the next one.