Build Real-World Apps with DeepSeek V3: 98% Cheaper & Better Than GPT

Key Takeaways

in this video I'll introduce you to what might be the most exciting open-source model currently available deep seek V3 it's the first open-source AI model to outperform well-known closed Source Solutions like GPT 4 or clae 3.5 in various benchmarks that's not possible I'll show you how to access it how well its chat feature Works whether it can keep up in terms of multimodal capabilities and how to integrate it seamlessly into your own projects to do this we'll set up a professional chat solution in just a few minutes which you can use as a foundation for your own projects why make the switch when you compare the price performance ratio of deep seek to GPT or Claude there's a clear winner what's the catch deep seek is based in China so it's unclear how your data might be used this means it's especially suitable for projects where data privacy isn't the main concern okay let's head over to the Deep seek homepage and find out just how good this model really is first we'll take a look at the benchmarks deep seek V3 outperforms Claude 3.5 and GPT 4 on benchmarks like mlu and drop these tests measure how well a model handles a wide range of academic tasks and complex text reasoning overall deep seek V3 shows it can compete with and often surpass top closed Source models so let's start by trying out the Deep seek chat for this we first need to log in I'm going to use my Gmail account for that all right now we've got a chat interface let's start with something simple tell a joke we can see deep seek get to work and tell a joke I've seen quite often before now we want to ask something more challenging namely how many R letters are in Strawberry this is a classic question that many other open source models struggle with but it looks good on the first try it answered correctly next I'd like to try the analysis of an image so I'll upload a receipt of Walmart this receipt lists multiple items and I'd like to find the price of one in particular the pet toy at the top then I'll directly ask a question about this item specifically how much this pet toy costs looks good the model correctly States 197 that means it analyzed the receipt correctly now let's try something more complex we'll take the classic where's Waldo image and ask where Waldo is but that doesn't look good it says the image has no visual clues about Waldo's location which of course isn't accurate all right but now we actually want to work with the API and I want to show you that it's fairly simple and nothing to be afraid of when we log in we first see a note that current prices are still lower but will soon be raised let's see what that means and what the current costs are so we're using the model deep seek chat and right now we pay4 cents for 1 million tokens input and 28 cents for 1 million tokens output but we can already see the price will increase soon it's relatively quick to add money yesterday I added $2 for testing and it was very easy via PayPal or credit card you can see that even though I've tested it thoroughly I've hardly used any credit we can already create an API key which we'll definitely need later and should keep somewhere safe we give it a name and click on create API key then we copy the key it stor it somewhere now let's see how to make our first API call we'll go to python for this and we can see that the code is structured pretty simply and looks a lot like the open AI API integration we copy that code and create a new project by first making a folder called Deep seek chatbot we switch to that folder and open our editor of choice in my case cursor the first thing we do here is create a virtual environment to keep libraries separate we activate the environment and create a new file called app.py then we open it paste in the copied content and see that open AI is underlined that means we still need to install the open AI package we do that once with Pip install then we have to insert our API key and now we already have everything we need to run the script and try a test run we see that after sending hello we get back how can I assist you today that's great so I'll make a git commit of the current state and to do that I run git in it I'll add the venv directory to the G inor so it won't be added and then I'll commit the current status all right next we want to stream the response that means as soon as the language model starts generating text we'll output it immediately for this we wrap all the logic in a method called stream response and we tell the language model to stream the response our goal is for this stream response method to be called from an endpoint therefore we return each Chunk in the response as soon as it arrives we can use the yield method for this that handles each new chunk as it arrives now it's time for the flask server for this let's create a new python file server.py in server. pii we first import flask request and render template from flask and of course we need to install the flask library after we've done that we want to create two endpoints one endpoint returns a simple HTML page that displays the chat window and communicates with the second endpoint via HTTP streaming we'll create the index HTML page in a minute the stream endpoint will be accessed by the HTML page via vanilla JavaScript code we simply send the entered message as a parameter and call the stream response method from apppi we just need to tweak the stream endpoint slightly specifically we create an internal method called generate that calls the stream response and the result is then returned by the endpoint we can now actually start the server because we haven't built an HTML page yet we can test our endpoint via the specified Port by doing a simple curl request you see here I'm doing a quick request with the message hello how are you when we fire that off we nicely see how the result streams in so it really is delivered via HTTP streaming now of course we want a web page with a great chat interface that uses this endpoint we create a temp templates folder and inside it a file called index HTML there we can add a heading open it in the browser and it looks good it's being served here I paste some predefined code which is purely HTML with JavaScript we have a few containers a chat container and a JavaScript method that calls the stream endpoint and puts the response into the appropri container this is just a basic example that can be used as a blueprint many of my clients use it to integrate the end points into their existing websites we can also check it out live we have a nice design and if I say hi we can see how the AI responds there are no limits to your styling options I went for an 80s arcade look but with tools like cursor you can change the design anytime and adapt it to your websites and projects