NEW Deepseek R1 Beats OpenAI o1!🤖 Start Using In n8n & Cursor Open Source AI Reasoning Benchmarks

So as a lot of you know deep seek has done something recently that many of us did not expect and we are seeing open source catch up maybe even surpass some of the frontier Clos Source models faster than we thought in 2025 just a couple weeks ago we covered deep seek version 3 and how it's very incredible you can do many cool things with it it's very good and and comparable to models like Sona 3.5 GPT 40 and recently we've seen

Open ai's 01 preview 01 01 Pro and 03 low tuned and high tuned get some crazy scores on the arc Benchmark now if you're not familiar with this test this is a test that basically shows from 2019 with gpt2 2020 with gpt3 2021 2022 we're seeing slow slow progress and then boom 2023 we're starting to get a little bit more with GPT 4 2024 we got GPT 40 right here so just a few months back we're still only getting about 5

To 10% on this actual Benchmark this is the AGI Benchmark and then boom the new paradigm happened with 01 preview with the new test time compute Paradigm and then as you can see we're really hitting an exponential curve of growth right here with 01 getting over 25% 01 Pro over almost 50% and then the unreleased 03 low tuned and high tuned getting closer to 90% re sit in about 87% with the high tune model so if you don't know the ark

AGI test evaluates whether an AI system can autonomously acquire new skills and solve novel tasks without human intervention serving as an early indicator of artificial general intelligence and this is everyone's Northstar right now especially open AI but now we are seeing some crazy benchmarks with deep seeks new R1 model and as you can see here like we say never always you know believe 100% in benchmarks especially when the company is putting out their own benchmarks but you can get a good

Idea and we're going to do some tests in just a second here but you can see deep seek R1 deep open ai1 right here deep seek R1 32b open AI 01 mini and then deep seek version 3 we have different types of tests right here code fores math and as you can see deep seek is on par if not even surpassing 01 in certain areas all of that with about a 80 to 90% cost reduction opposed to using open ai's 01

Model so all of this is crazy news for the open- source Community we're going to dive into this model run some tests and I'm even going to show you how you can use it today in a practical use case whether using n8n cursor whatever the case may be let's dive right into it all right guys so before we get into it I just want to give you a quick recap I'll leave this link to this thread down below there's many people

Posting about this on Twitter X so as you can see it's been this was posted a couple days ago it's been over uh 80 hours now since deeps gar1 was released open source you can see many different use cases some people are considering this as AGI at home running deep SE gar one on 7 M4 Mac minis right here we got deep seek R1 1.5b running fully locally in your browser by web GPU so many different ways you can use this

Someone right here building a perplexity clone with deeps R1 in about an hour without writing a single L of code deep seek R1 creating a rotating triangle with a red ball in it and 01 could not do this building a rag app to chat with your documents using deep seek R1 running loc this is another cool thing guys and I'll show you this in just a second deep seek also released smaller fine-tune models like llama 8B achieving similar or better performance

Than GPT 40 or Claude Sonet you can also run deeps seek R1 distilled quen 1.5b 100% locally in your browser really with anything but here it's with web GPU again and then a really cool thing is the Deep seek Chain of Thought where you can actually read this out loud and it literally sounds like a human thinking it's really cool to actually see this and as we continue with this and this whole test time compute new paradigm I believe that we're

Going to see more and more of these AI models acting and thinking as an actual human would through these problems and another tip is you can actually make any model smarter just by extracting the reasoning from Deep seek Reasoner as you can see here by the way just extracted the reasoning from Deep seek Reasoner and he used it on GPT 3.5 to turn it into an absolute genius this is something that open AI has been so secr about with in terms

Of how this their models 01 actually go about their thinking process and now we can actually see behind the curtains with deep seek so this is really powerful so guys if you haven't already tried this out I would definitely recommend going to deep seek.com all links that I cover in today's video will be linked down below and giving this a try so I'm just going to show you an example right off the bat because this is just so cool now you

Can create a free deep seek account and you can also get the app on your phone so I have the app on my phone too and it's pretty awesome it's it's actually a nice user interface the one thing is the caveat is this is a Chinese company um so do with that what you may I'm sure you already heard now about project Stargate but this is huge because USA is really trying to take the lead in terms of the AI race

And you know now we're seeing an open source Chinese model that's basically on par with the top Frontier Model that is closed Source at the moment so it will get interesting for sure and keep in mind they did all of this at a fraction of the cost of What it Took open a ey to actually achieve the same result so pretty interesting stuff okay so you can ask it sometimes very basic questions and it will do a large thinking process depending

On the question or you can ask it a little bit more intricate questions give it a little bit more context um but here I'm just asking how is deep SE R1 a different Paradigm with test time compute how is it different from other AI models and also how is it different and similar to open AI L1 and L3 strengths weaknesses I'm doing a YouTube video on this topic so we can turn on deep seek R1 right here the Deep think you

Could also toggle on the search or toggle this off I'm just going to keep it on for now now this will search a lot of different websites that are in Chinese so if you ever do get a Chinese response sometimes you may get that you can just say hey tell me in English now if you have search on you won't be able to upload a file but if you don't have it on then you could actually go ahead upload a dock

Or an image let's go ahead and click send as you can see it is started its thinking process so it's all Al first searching the web right here so 50 resources were found as you can see you can scroll through these see all 50 of them some are in Chinese some are in English and you can see that it's actually thinking so it thought for 10 seconds and I'm not going to read the whole entire thought process but I will read

You a bit just so you get an idea of how cool this is okay I need to explain deep seek R1 and how it uses a different Paradigm with test time compute compared to other AI models and compare it to opening eyes 01 and 03 let me start by understanding what test time compute means from the search results test time compute refers to the computational resources used during inference not during pre-training this is a shift from focusing solely on pre-training with

Massive data to optimize how models stin during actual use looking at page one it mentions that deep seek R1 and Kimmy K 1.5 are part of this test time compute Paradigm the idea here blah blah blah as you can see it's going through the different resources here and we get our answers so test time compute Paradigm how deep seek R1 differs so test time compute refers to blah blah blah it's just explaining that how it thinks step by step throughout the

Problems enabling for better reasoning and accuracy without solely relying on scaling pre-training data so a few Innovations with R1 is reinforcement learning first approach so unlike traditional models that rely on supervised fine-tuning sft to Prime reasoning r10 r1's precursor applied R1 directly to the base model mimicking Alpha Zero's trial and error learning this allowed the model to develop reasoning behaviors like self-verification and reflection autonomously minimals supervised data so R1 later combined RL reinforcement learning with a small cold start data set

Of longchain of thought examples to improve readability and stability avoiding the need for massive labeled data sets efficient reward design instead of training complex reward models R1 uses accuracy example code execution results and format rewards example structured output to guide reinforcement learning reducing computational overhead and distillation for scalability so r1's reasoning capabilities are distilled into smaller models like 1.5b to 70b parameters enabling cost effective deployment while maintaining performance close to larger models all right now it's going through the strengths and

Weaknesses between deep seek R1 and open AI different contacts architecture the cost right here so it's 96% cheaper through the API the openness of the model the performance the training efficiency and then it talks about 03 the Unseen competitor strategic implications takeaways for your YouTube videos so visual comparison cost analysis right here case study broader contact so it is really really powerful as you can see we got a link to the GitHub repo right here or deep seek and like I

Mentioned before we have deep seek V3 we have deep seek V3 base r10 R1 distilled quen 7B quen 7 quen 14b llama 8B quen 1.5b quen 32b llama 70b and deep seek R1 so you could actually use these models locally now you would want to use the distilled versions if you're running it on your main computer most likely unless you have some sort of supercomputer or you're running it on some GPU as you can see we have the 671b model which

Is 404 GB then we have the 70b one 32b 14b 8B 7B 1.5b so you could probably run the 14b depending on your computer you would simply just go to O Lama you could also use this on a hugging face or whatnot um but if you're using o llama you could simply just run this command in your terminal if you have o llama installed and then you could get this model up and running locally you can also use LM Studio or

Any other alternative of course for this now I'm going to do a lot more videos on deep seek R1 and how to actually use this in real applicable use cases but I want to show you two main ways if you don't already know just a few quick start actions that you can take to start using this right away so first things first like I mentioned before you can check out the deepsec website if you go to the Deep seek API platform

You'll be able to see your usage top up billing and you can create a new API key you can go view their docs over here and you'll see that they provide a base URL right here as well as you would just input your API key and then over here you'll see the model name so if you want deep seek chat that will be V3 and then deep seek Reasoner right here is R1 you can see the context length Max tokens Max

Output tokens 1 million tokens input price so 14 cents and this is with cash hit and then 1 million tokens input price with cash Miss 55 1 million tokens output price 2.9 19 now unfortunately we have a few unsupported features right now such as function calling Json output and fim now there are a few workarounds and I'll cover some of these in future videos as well but let me quickly show you how you can actually use this in something like NN

And cursor So within NN I just have a very simple template right here as you can see this is just a agent right here and you there's really two main ways you can go ahead and use this so the first way right here as you can see it is Jeep seek open router so if I click on this um right here so this is just an open AI chat model right here all right so we're going to connect this here now

If we open this up you'll see this is my open router connection so I have my API key posted right here and then the base URL which is open router. a/ API slv1 so if we go to open router here you can see that we have deep seek R1 I'll leave a link to this down below you can see it also mentions some additional providers of deep seek R1 so you could use deep seek which I'm going to show you in

Just a second you could also use deep infra to gather Ai and fireworks and you can see the latency the cost and everything comparison right there we're just going to go ahead and copy this model right here so it's deep seek SL deep seek R1 and once you have your base URL and API posted right here and you have your connection made you can simply just go to the model right here and you're not going to use fixed you're going to

Use expression and you're going to paste in the model ID right here once you do that you really will have deep SE in your agent right here we could go ahead and test it out right now so I'm saying hi what model is this and as you can see it says I'm deep seek R1 next we have another agent right here and this is the exact same except the connection is just directly through deep seek so if we go to our

Connection you can see the base URL is this api. deep seek.com we got our API key right here from deepsec and the model is deep seek D Reasoner you can get this from the API docs as well but just simply put that in and you should be good to go so now if we ask what model is this again and boom we get deep seek right here now you can do the exact same thing in cursor so to do this you

Would simply just go up to your settings Cog right here and you would go to your models once you're in models you would want to enable your open AI API key right here and you would want to override the open AI base URL and just do open router. a/ API slv1 p in your API key and then you want to add a new model right here and add that same model ID which is deep seek de seek R1 and then as

You can see we asked it what model this is and it's saying deep seek R1 now it would be cool if you could actually see the thinking process within cursor maybe if there was a toggle of some sorts but unfortunately you can't do that same thing with NN you can't see the thinking API request at the moment and as you can see we get our answer right here going over what deep seek R1 even is but other than that guys that's

Prettyy much it for this video I just wanted to give you a highlevel overview as well as some practical use cases that you can start using deep seek R1 right now after watching this video in nadn and or cursor now if you want me to cover any specific type of video surrounding deep seek R1 showing you how to build certain things like AI agents or whatever the case may be maybe an n8n or an actual code let me know in the

Comments down below I'm definitely going to be doing a lot more AI agent videos coming soon make sure to subscribe like the video and comment down below to stay up to date for that if you're new to the channel we upload videos all the time on AI agents AI news whatever the case may be so subscribe to stay up to date for those videos and then also two join our free Facebook group and Discord channel to connect with our community as

Well as to get behind thescene resources and templates and tools so I leave a link to that strid community.com and then definitely check out our stride AI Academy it is an exclusive community that we're building and it's going to be jam packed with loads of free values so make sure to join that while it is still free also too guys if you run a business and you need help implementing AI agents like Ai call center AI appointment center naden agents whatever

The case may be into your business whether you're a marketing agency a SAS company service-based business B2B whatever the case is book a call down below at executiv stride. comom slapp and we can see if it's a fit or not other than that guys let me know in the comments down below what you think about deep seek R1 is open source going to catch up to close source and surpass them and stay tuned for more videos on cursor NN and a

Bunch of stuff that I have planned coming soon other than that guys keep hustling keep grinding and of course guys accelerate your stride take care