[Music] Hi, welcome to another video. So recently, multiple things happened from the GPT-4 mini launch to a Windows crash, and one thing that got buried by all this was that the DeepSeek V2 chat model got an update. If you don't know, then let me tell you that DeepSeek V2 was launched about a month or two ago, and it was really great in both benchmarks and real-world usage. It was mainly focused on all types of general usage. Shortly after that release, they also released the DeepSeek Coder V2 model for coding, which was also pretty good. But now they have released an update to their DeepSeek V2 chat model, which makes it even better than before. Just like the previous iteration, this is a Mixture of Experts 236 billion parameter model with 21 billion active parameters. Although there's also a light version of it, which has only 16 billion parameters, it doesn't have any new updates. Anyway, this new iteration of the model is also ranking really well on general leaderboards as well as the

coding Arena leaderboard, so that's also cool. If we talk about the major improvements within the model, you can see that in the Human Eval Benchmark, it has gained about 3.7 points. In the Math Benchmark, it has gained about 17 points. In the BBH Benchmark, it has gained 3.7 points. In the IAL, it has gained about 13.8 points. And in the Arena Hard Benchmark, it has gained about 26.7 points, which is the biggest gain. They also say that the instructor action following capability in the system area has been optimized significantly, enhancing the user experience for immersive translation Rag and other tasks. So overall, it has gotten some good upgrades in performance across almost all sectors, which is obviously quite cool. The previous benchmarks of this model were already quite good, but now it's even better. The model is available on Hugging Face. Although Alama currently doesn't have the updated model yet, the Deep Seek Chat model on their site is also updated to use this new model.

let's test it and check it out. But before we do that, I'd like to tell you about today's sponsor, On Demand. Creating AI products can be hard, but with On Demand, it doesn't have to be. On Demand accelerates AI development with advanced rag systems, slashing your development time. With On Demand, you can access a wide range of models from OpenAI, Gemini, and more. You can integrate these with multiple plugins from the marketplace, giving your LLMs real-time access to various services and knowledge bases. You can customize them even further by creating your own plugins from your knowledge base or API, tailoring the LLMs to your specific needs. You can also share your plugins on the marketplace for rewards. From aviation and event experts to real estate and home automation, On Demand lets you create AI agents for every need. Visit On Demand and get started now for free. Now let's get back to the video. To use it, you can go to their site, register for an account, and then use it for free. Now here are the nine questions that I'm

going to try it against. So let's get started. The first question is: What is the capital city of the country whose name ends with Leah? I'm referring to the country name here. Let's see if it can answer it. Okay, here's the answer. And as you can see, it answers it pretty easily and accurately. So let's mark this a pass. The next question is: A baker made 240 cookies. He packs them into boxes, each containing 12 cookies. How many boxes does he need to pack all the cookies? The answer to this should be 20. So let's check if it can answer it or not. Okay, here's the answer. And it answers it correctly. So let's mark this a pass as well. The next question is: Lucy has twice as many candies as Mike. If Mike has seven candies, how many candies does Lucy have? The answer to this should be 14. So let's send it. Okay, it answers this correctly as well. So that's also a pass. The next question is: If a regular hexagon has a short diagonal of 64, what

is its long diagonal. The answer to this should be 73.9. So let's check its answer. Okay, so it doesn't answer it correctly. So let's mark this a fail. Now the next questions are coding-based. The first question in this category is: create an HTML page with a button that explodes confetti when you click it. You can use CSS and JS as well. Let's send it and check. So here's the code. Let's copy and preview it. Okay, so this works and looks pretty fine. So this one's also a pass. The next question is: create a Python program that prints the next X leap years based on user input. Let's see if it can give a working code or not. Okay, here's the code. Let's run it and check. Okay, so this one is also working fine. So this one's also a pass. The next question is: generate the SVG code for a square face. Let's see if it can generate it or not. Okay, so it has now generated the code. Let's preview it. Okay, this looks pretty cool. So this

one's also a pass. The next question is: create a landing page for an AI company. The landing page should have four sections: header, banner, features, and contact us. Make sure that the landing page looks sleek and modern. Let's see if it can generate it or not. Okay, so here's the code, and let's preview it. So, this looks fine and has all the sections we wanted and everything, so this one's also a pass. The next question is: write a snake game in Python that works on the terminal. Okay, so here's the code that it has generated. Let's run it. Okay, so this one's also working correctly. It works pretty well and has the basic stuff we'd want, so this one's also a pass. Now, here's the final chart. You can see that it almost passed every question except for one. This model is literally very cool. I mean, it's on par with all the greatest models, and the best part that I really like about it is that it's open source. And if you use their chat platform, then it's also fully free to

use with no hard limits. Also, if you think that GPT-3.5, that OpenAI recently launched, is cheaper, then I think you don't know that this model's API is even cheaper than GPT-3.5, while providing extremely good quality. This DeepSeek model only costs 14 cents for input and 28 cents for output, which is obviously very cool. I mean, it's crazy how people on the internet keep praising the GPT-3.5 model for their cost, whereas a better model with better quality is even cheaper than it. So, overall, the DeepSeek V2 chat model was really good, and now it's even better. I really like how well DeepSeek's and Quinn's models are performing. Overall, it's pretty cool. Anyway, let me know your thoughts in the comments. If you liked this video, consider donating to my channel through the Super Thanks option below. Also, give this video a thumbs up and subscribe to my channel. I'll see you in the next video. Till then, bye. [Music]

DeepSeek-Chat-V2.1 (0628) : The DeepSeek-V2 LLM GOT EVEN BETTER! (Fully Tested & Better than GPT-4O)

Facing the future with AI! 

Here you'll find content about multiple AI Tools that are actually useful and sometimes free.