Transcription
So Deepseek has recently introduced Deepseek R1 version 2, and I'm already testing it for different kinds of coding tasks. For example, this particular one where I just built this interactive Rubik's cube simulation. The only problem is this demo is not working. There is a working version as well, which is this one, which is not working as it should, but somehow it is like, kind of, it scrambles, after that it tries to like solve, but still there are some mistakes.
So this particular version is without Deepseek R1, without Deep Think. So this is a Deepseek R1 version 2, but without Deep Think. This particular one is Deepseek R1 with Deep Think.
Then I just created this basic 3D runner game. It's a shame that I have to like call it a game because, being a game developer myself, I just cannot call it a game. But these are things that I do for YouTube. But anyways, um, so yeah, this is just like a very basic, horrible 3D game. But anyways, the prompt was very simple, just create a 3D running game. Then it had its thinking for 17 seconds, and after thinking for 17 seconds, it created this very basic HTML code and then created this very basic 3D runner game.
Similarly, I just created this basic landing page using DFCarvin, and I was kind of like impressed by the first version because the prompt that I gave was very simple. It was like, just create a SAS landing page. Then it had a thinking for like 11 seconds, and after that it just like started writing code. Then there were a couple of changings, and then just this is the finished version. So this looks kind of like really professional so far. But yeah, I mean, it's just like a standard SAS landing page.
Then I also tested it inside client. So I'm actually using Deepseek R1 version 2 inside client through API key from Open Router. If you don't know about that, so you can just like go to Open Router. Open Router is a platform where you can actually get all of these different like models. You can access all of these different models through API key.
So this is Deepseek R1. This model costs $0.50 per million input tokens and $2.18 per million output tokens. There is a free version as well. You can access that as well. So this is the free version, Deepseek R10528. This is the free version, and it costs $0 per million input tokens and $0 per million output tokens. You can start using it inside Clone or any other AI coding agent, maybe Rule Code, or maybe any other AI coding agent. You can start using it from here, and this is absolutely for free.
But in case if you want to like use it for your actual project, so I highly, highly recommend you like using this version, which is, which costs you around $0.50 per million input tokens and around $2.18 per million output tokens. By the way, Deepseek R1, this particular version is around 2.5 times more costly than Deep Seek version 3, which is again a good enough model, which was launched I think in March. And Deep Seek version 3 is a 658 billion parameters model, and it's a mixture of expert model. But this Deepseek R1 is comparatively better in terms of like thinking, in terms of overall reasoning, and of course for coding tasks as well.
Context is 163,000, which is kind of not enough when it comes like using it on a large code basis. It is good enough like basic HTML codes, JavaScript code, but when it comes like using it inside like a proper project where there's like thousands and thousands of lines of codes, so it might not be enough.
Anyways, this is a 671 billion parameters model with 37 billion active in an inference pass. It's a fully open source model, and it actually competes with a lot of other these famous models, something like OpenAI3, Gemini 2.5 Pro, and Quen 3 235B, and Deepseek R1, of course.
So this is a benchmark comparison. If I compare it with all of these different famous models, so in this graph, this is Deepseek R1, this light blue one, and then this one is Deepseek R1 version two. So this is the old Deepseek R1. This is our Deepseek R1 version two. And you can see Deepseek R1 version 2 beats almost all of these different models in all of these different benchmarks, including this live code bench and humanities last exam.
So, so the most shocking thing for me is this model is beating Gemini 2.5 Pro in some benchmarks. So that means this model is good enough.
Now, if you're interested in testing this model inside client or root code, you can simply just go to V Studio and make sure that you have installed client and just go to settings. From here, you can just like click on this API provider. From here, just choose Open Router, and from here you can just like provide this Open Router API key. As I told you earlier, you can simply go to Open Router, sign up. You are not required to attach any sort of credit card if you're not using any of the paid models. So you can just like use this free one. This free model, which costs $0 per million input tokens and $0 per million output tokens. You can then create an API key, and then after that, you can simply just provide that API key in here and then choose Deepseek R10528 from this models drop-down section.
Then you can actually provide custom instructions for this model if you want. The only problem is this model does not support images. It does not support browser use, and it does not support prompt caching. For me, the big downpoint is this model does not support images. Most of the times whenever I'm working on any sort of like project and like doing web coding, so like providing images to my AI code editor actually helps me to explain my scenario in a better way. But this is actually a drawback that it does not support images.
But if you still want to use it, maybe like for code analysis or for any other coding task, you can use it, of course. Also, you can use it for free inside chat.deepseek.com, deepseek.com, and you can use Deep Think as well from here. And you can see, let's suppose if you want like create any sort of like landing page or any sort of like tool, or make this kind of like basic 3D game, it can write code for you. And of course, it can then run that code for you. So let's suppose if it just like wrote this code for you, and then from here you can simply click on this run button, and then it can run this code for you inside directly inside your browser.