Speaker 1: OpenAI's O1 model kind of took over the world, especially with its ability to think, as we now call it these reasoning models. Well, we finally have an open source equivalent. This is coming out of a Chinese AI research lab called DeepSeek. They have released DeepSeek R1, which is an open source reasoning model, which is equivalent to OpenAI's O1 model. And as you can imagine, the whole AI world has gone crazy over this because it is equivalent to OpenAI's O1 model, and especially for those people who pay $200 a month, this is equivalent to that and it is free. And in today's video, we are going to be looking at how you can use DeepSeek R1. You can run it locally. And if you don't want to run it locally, or if you don't have the necessary hardware to run it locally, we'll also look at how you can access it for free. So before doing that, let's take a quick look at some of the comparisons of DeepSeek R1 with OpenAI. So we have both AIME, this is a math benchmark, CodeForce. For coding and math, it has done exceptionally well. And in certain cases, it is above, for example, Math500, DeepSeek R1 is above OpenAI's O1 model. Obviously, this compares it to O1 model, not the O3 model, which is going to be released in the next couple of weeks, as Sam Altman tweeted. Anyway, if you want to run DeepSeek R1, so first of all, we are going to go to OLAMA and we are going to download and run OLAMA. So head over to olama.com. And if you guys want a detailed guide, I have written a detailed guide on how to run LLMs locally using OLAMA. So this entire article, I will be putting it in the description box below. So you can go ahead and take a look at it there, especially the memory requirements. Anyway, so once you have downloaded OLAMA, head over to the model section. Here we're going to see DeepSeek R1, click on that. DeepSeek R1 comes in various formats, various parameter count. So we have the 1.5 billion, 7 billion, 8 billion, 14 billion, 32 billion, 70 billion, and the full format, the full 671 billion parameter model. And obviously, this is not something you're going to be able to run locally. So we have here, these are the parameter count of what the models are. We have 1.5 billion, this QUEN and OLAMA, this UC is basically the different type of underlying architecture that is used. So in this video, I'm going to be using the 8B LLAMA DeepSeek R1. So this is the command we are going to run. For the 1.5 billion parameter model, you would need around 8 GB memory, 7 billion and 8 billion, you would need around 16 GB memory, 14 billion, you would need around 32 GB memory. And for 32 billion, you would need around at least 64 billion, but to run it, 64 GB memory. Or to run it comfortably, you would need around 128 GB memory. So if you have those, look at your hardware, what kind of hardware you're running, and if you have the required hardware, you can just pick whatever model you want. Specifically, if you want to run the equivalent of the O1 model locally on your laptop, I would suggest getting the 7TB LLAMA DeepSeek R1. You would need at least 128 GB memory to run this locally, and obviously we have the full 671 billion. I don't think you would have the hardware. If you have the hardware, congrats, that's a very big setup. Anyway, so we are going to be running this model. I'm just simply going to copy this, and I'm going to head to my terminal. So here I am. Make sure that LLAMA is open, so LLAMA is running. So I have LLAMA running here, and I'm going to run the command, LLAMA run DeepSeek R1. So this is going to download the model here, which is, if I take a look at it, so that 82 GB, it is 4.9 GB file size. So it's going to take some time depending on your internet speed. So let's just wait for it to download. It is almost downloading, so it's almost done. So it says success, that means it has successfully downloaded the model. Just waiting for it to... Yep, so when we see these three arrows, that means you can now type in any message. So this is the reasoning model, so it's going to think, and we can see its thinking process. One of the very big questions that people ask is how many R's are in a strawberry. Instead, I'm just going to say, why do people ask how many R's are in the word strawberry? Yep, let's ask this question. So here we can see in these arrows that it is thinking, so this is its thinking. This is not the final output, this is just its thinking process. So it's saying, okay, I'm trying to figure out why people ask how many R's are in the word strawberry, and you can see, depending on your memory, this can be fast or slow. Keep in mind that I downloaded the 8 billion parameter model, and I'm running this on MacBook Pro with 16 GB memory. And if you have a higher one, you can go for an even more powerful model, and if you don't have this, you can go for a lower one, but again, the quality might not be that great if you go for a lower one. So while it is typing, it's doing its thinking, let me just show you how you can run this for free if you don't want to run it locally. So for that, we would have to go to chat.deepseek.com. So this is basically the chat.gpt.com equivalent from DeepSeek, so we can ask the same question here, why do people ask how many R's are in the word strawberry. So we're going to click DeepThink, so this is when it starts using its DeepSeek R1, the reasoning model. So as you can already see, it is much faster than me running it locally, so this is completely free, you can go ahead and do it, but keep in mind that based on their terms of service, whatever you type in here, they will be using it to train their model, so do not put anything, any sensitive information here. So we can see here that it thought for 19 seconds, and it gave out, so first of all, it did correct, so it did the three R's, there are three R's in it, and some answers. So let's go back to our own running locally, so this is where the thinking process ended. So this entire section is its thinking process, and after thinking, this is the final output which we get here, so they're always interesting. So as you can see here, this is basically how you can run it locally, you can ask any question, and obviously, if you're running it locally, that means it is completely offline, you do not need an internet connection, while if you go to chat.deepseek.com, you will need an internet connection. So that's it. Now another important thing which I find very interesting, especially on running it on their website, is that you can run at least HTML directly onto their site. Since I have a website where I make courses on machine learning, let's say if it can create a landing page for me, so can you create a landing page, if I can move this, yes, can you create me a landing page for a website that sells, teaches machine learning courses. So let's run it. So it is doing its entire thinking process, it's very similar to what O1 does, but keep in mind that O1 has a 50 prompt limit per week, while this is completely free, and I'm pretty sure there is going to be some limit for this, but they don't say it, but if you're running it locally, you can ask as many questions as you want. So it just gives me this, I think I should have been more specific, asking for code. So I'm going to say, can you give me the code for this. So it's running. While it's running, we currently have the state-of-the-art reasoning model that is from OpenAI's O1, and this is equivalent to that. OpenAI did mention that they are going to be launching their O3 series model in the next couple of weeks, so that's going to be very interesting. If you guys are interested in that, make sure to give a follow, see when that comes out, and my review on that. So it is currently doing this, generating the code. So here, once it's done, we can simply run the HTML here. I think it did a pretty decent job compared to considering that it's completely free. So yeah, that's pretty much it. So this was how you can run DeepSeek R1 locally on your laptop, and also how you can use it for free. And as always, thank you for watching this, and I'll see you next time.