DeepSeek R1 0528 in 6 Minutes

--- type: transcript date: 2025-05-29 youtube_id: 8k1ul7m4IyQ --- # Transcript: DeepSeek R1 0528 in 6 Minutes In this video, I'm going to be taking a look at Deepseek R10528, which was just released a number of hours ago. Now, what's interesting with this release, they didn't release any model card. There was no official announcement on X or any other platform, but there are a number of really interesting things coming out about the model. In this video, I'll go over what we know about the model, as well as point you in the direction on how you can get started with the model. And then, of course, I'll also go through some demonstrations, whether you want to use the API or if you want to have a chat interface. First thing right off the bat, the model does have an MIT license. So you can feel free to use this build on top of it. If you want to further post-train this and use it in other ways that you want to leverage, you can go ahead and do that. First up, in terms of some of the benchmarks that we do know about on live codebench, this sits just between 03 mini on high mode and just shy of 04 mini on medium mode. Now, if we just take a look at the chart here, the big difference with these benchmarks, while it is sitting in fourth place, if we look at all of the other models, with the exception of Quen, all of those are closed source models. Quen and Deepseek are the only options in terms of performance for open- source options that are out there right now. Next up, just a number of minutes ago on the local llama subreddit, someone posted that this model scores the same as Claude Opus on the Adar Polyglot benchmark at 70.7. The old R1 was 56.9. So, if I just take a look at the Adar Polyglot benchmark, we can see all of the latest Frontier models here. In terms of the reported results, we have Claude Opus with No Think at 70.7. But the cost is going to be the one big difference as well with this model because the one thing with open source models is there's going to be a ton of competition in terms of where to host this model. And there are even some free options that are available that I'll show you a little bit later in the video. Now, in terms of the now infamous hexagon with the ball bouncing in it, here is the example of the latest DeepS R1 model, and it doesn't look like it is the best in terms of the physics. So there are definitely some better physics in terms of how the ball bounces up and down that I have seen with some other models. But just a fun little example, another benchmark that I saw out there was the extended New York Times connection benchmark which has it jumping from 38.6 to 49.8. And with this model we have it sitting just shy of claw for opus thinking with 16,000 thinking tokens. So just above Gemini 2.5 Pro on this particular metric. One quick note with the model is it is already available on open router. If you are interested in trying this out from the API or from a chat interface for that matter, you can try this at open router. Right now, at time of recording, there are six different providers just hours after the release. I'd imagine this is going to probably increase substantially over the next few days as different model providers begin to host this. What's great with open router is you can see things like the median throughput for the different endpoints as well as of course things like the cost for the input and output tokens. Now one thing to know with this model is it is incredibly cheap when compared to some of the other models in terms of the context window as well as things like the throughput as well as the time to first token. All of the providers are going to vary a little bit in terms of what they allow for context window as well as some of the other performance benchmarks. But just to put this into perspective, it's 50 cents per million tokens of input and $218 per million tokens of output. That puts it around a dollar and something for the blended rate. Just to put that into perspective, in terms of cost, it's going to sit somewhere between Gemini 2.5 Flash Reasoning as well as 04 mini high in terms of the blended rate. Definitely considering the capability of the model, this definitely is a very competitively priced model. And I wouldn't be surprised if we begin to see other price adjustments over the coming days as all of these different models begin to try and compete and win everyone's business, especially if there begins to be some momentum around the model like R1. Now, in terms of being able to access this, you can try this out on chat.deseek.com. And what I'm going to say within this is I'm going to say build a beautiful landing page for my brand developers digest with modern styles, subtle animations, as well as compelling copy for a technical content creator. If you haven't used an R1 model before, what you're going to see as the first step is the thinking trace here. We are building a landing page for a developer digest, a brand for a technical content creator. The page should have modern styles, subtle animations, so on and so forth. It's going through and breaking down the task within here. It's building out an HTML page for us. One thing with this that is interesting is it is generating a lot of CSS. Now, one thing that I have noticed with a lot of other models is oftent times the models do generally gravitate towards something like Tailwind where it will have all of these different classes biases towards using these CSS frameworks rather than actually writing out CSS. It is interesting right off the bat to see that it is writing out that styling by scratch. The nice thing with the DeepSeek app is I can go and I can run things like the HTML. And here we go. This is what it generated for us. By the looks of it, it is a very competent model in terms of UI development. There definitely are a ton of little nitpicks here. There is the Twitter bird, for instance, with an underline as well as this underline for the GitHub icon. And if I just go through the page here, it definitely looks pretty good, but we do have things like the quotes that are cutting off pieces of the text. But overall, in terms of the first pass with just a sentence of instructions, arguably it did pretty good. We even have this interesting artistic background with these different shapes. We also have these subtle hover effects here. Now, one last thing that I did want to call out is there is currently a free endpoint on Open Router. You can try this out within the chat interface here if you want to try out some of these preset questions or put in whatever you can do that. In terms of trying this out, you can create an API key and then you can choose whether you have the SDK like the OpenAI SDK that you want to leverage to try out this endpoint completely for free. So, overall, that's pretty much it for this video. I wouldn't be surprised if we do see a model card coming out over the coming days. If and when that does get posted, I'll link that within the description of the video. But otherwise, that's pretty much it for this video. If you found this video useful, please comment, share, and subscribe. Otherwise, until the next