Kimi K3 amb Kimi Code: instal·lació i prova davant Claude i Codex

Major sign that the US may be losing its AI advantage to China. >> China is beginning to engage in what I'll call AI dumping. About a third of corporations are supposedly using Chinese lightweight open-weight models that [music] are cheaper. If I were she, I would just dump cheap AI into the US market. I think the US market crashes and we're immediately in a recession. [music] >> The open models are not as far behind as we thought. A new Chinese model just beat Claude. >> [music] >> It honestly might be over for Claude and ChatGPT. Kimiko free was just released and this changes everything we know about the Chinese open-source models. Because just look at the benchmarks. Kimiko free is right up there with Fable 5 and GPT 5.6, which is absolutely insane considering it's an open-source model. And you can even see in many of the benchmarks it actually scores over GPT 5.6. And in [music] the benchmark program bench, it actually ranks higher than both GPT and Fable. Same in this benchmark right here as well. However, you should know by now that the benchmarks don't tell the entire truth. So, is Kimiko free actually as good as these benchmarks say it is? The results might surprise you. If you don't know me, my name is Albert. I didn't start out technical, but because I learned how to use AI, I've built two million-dollar companies. Not because I'm the smartest, I just learned how to use AI. And now we're making AI education free for everyone. So, if you haven't already, make sure to subscribe. By the end of this video, you'll know if Kimiko free is actually as good as people say it is and if we need to be canceling our Claude plans cuz it honestly might seem like we have to.

Let's not waste any more time. Let's get into it. We have already taken a look at the benchmarks, which looks absolutely insane. I honestly couldn't believe this when I saw them the first time because this is completely changing how we should look at open-source models. Let me explain why. Chimera-3 is almost a 3 trillion parameter model, which is by far the largest open-source model ever created. You can see Deep Seek right here, which is in second place, is a 1.6 trillion parameter model. Chimera-3 is almost twice as large. It's honestly insane some of the things that people have built with this. Someone even built a black hole. These are some of the coolest things that people have built, but we of course need to be doing our own tests, know if it's actually a model that we should be using. And we're going to be doing that in just a second. To test this model, we're going to be using Chimera Code, which is the absolute best for what we're trying to build in this video, which is going to be absolutely insane. To install it, we're just going to search for Chimera Code, click on the top link, and this gives us this code right here that we can just copy into our terminal. So, I'm going to go back into Visual Studio Code, and in the top I'm going to click terminal, new terminal, and I'm just going to paste this in to install Chimera. After it's installed, we need to restart our terminal, so we can just close this down, and then I'll open up a new terminal, and then we can write Chimera and hit enter. And now you can see we're inside of Chimera Code. We can just write {slash} login now, and then login with a Chimera account.

This will open up our browser, and here we can just authenticate. You can see login successful. And if we go back, you can see it says login. One of the most important things is of course the pricing of these coding tools. And as you probably know by now, Claude Pro is $20 a month. But one of the biggest bottlenecks with Claude in general is the usage limits, especially if you're using Fable, which is what's on the benchmarks, you're going to be running into rate limits almost instantly. And soon Fable will also only be available on the Max plan, which is $100 a month. But that shouldn't come as a surprise, Claude is expensive. When it comes to Codex, you need the $20 a month plan. But to be fair, with Codex you do get a lot of usage. However, with Chimera you can get started for $15 a month. So, I'll say Chimera wins there. To do a fair test, I'll create three folders. I'm going to create a a folder that I called Chimera, and I'm going to create another folder that I'm going to call Claude. And the last folder is going to become Codex. Those are the three models that we'll be testing in this video. I'm going to go out of this session now, and then I'm going to CD, basically go into the folder of Kimmy, and same here, CD into Claude, and same here, CD into Codex. And then we can just rename these to be Codex, to be Claude, and to be Kimmy. Now we have all of the sessions. Now of course have subscriptions to absolutely everything. So we can open up Kimmy code in this folder, Claude in this folder, and Codex in this folder. And the first thing we're going to do is that we're going to test out the Kimmy model, because I haven't even tested it yet before making this video.

Is it as good as people say it is? Well, let's find out. So I'm going to give it an improved prompt by saying, "Please build an improved prompt for building out a video game. I want a one-to-one replica of Pac-Man with all of the features that Pac-Man should have, and the styling, so we get that one-to-one replica. Please build out that improved prompt." Claude is then going to build out an improved prompt for us, like that. You can see it has all of the things that should be included in Pac-Man. So now we can just hit enter. I'm very excited to see what Kimmy can build from just this. And of course we also need to remember to set it to the right model. Right now it's set to Kimmy 2.7, but we need to write {slash} model, and then of course set it to K3. Run the prompt again. Of course it's important that we actually use the right model. I'm very excited to see what it builds. Great, it is now finished. Let's restart it. It looks like we have a fully functioning Pac-Man game. We even have the sound effects as well. Going to turn those off though. This looks very, very good. I'm very impressed with how well this turned out. This looks very impressive. What about all the small details like being able to go over like through the other side right here? It has basically built it perfect one-to-one. Very, very impressive. However, not the best at Pac-Man. Time to compare as well with Claude and Codex. Here we have the game that Claude built. Looks similar, however it is a little smaller, but it is also a fully functioning Pac-Man. So Claude definitely also did very well on that.

Very impressive. And for the last one, let's do Codex revealing finder. I'm going to do this one of as well. All right, here we have insert coin. So, this has done a little more in terms of that. I don't know if I like that. I didn't ask it to do this. Insert coin. All right. Oh. This doesn't look very good. The frames on this is very bad. You can see the the game is way more slow-paced. And we just went through a wall right here. Again, very buggy. So, Kimmy and Claude definitely wins that. I didn't actually expect that. I thought Codex would perform very well here as well. But, we are running into the walls. It's very laggy. The frames are very bad on this. Very interesting. I want to do another test, which is one of the use cases I use AI for the most, which is front-end design. So, let's try and give a prompt again. We tell all of the models to build out a clean website. Uh write an improved prompt for a very, very clean website for a dental dental business. It should include all of the the images as well. It should have a very high converting form at the bottom. Include all of the sections that a high converting website has. And please do a detailed description of the theme that should be very futuristic and very modern. Write an improved prompt for that. Code type will write out the improved prompt, so we can use that for every single model. This looks very, very good. So, let's give that to Kimmy. And let me copy this and give it to Claude as well, and to Codex as well. The exact same prompt. And let's see how they perform. All three models have now made a detailed plan of what they want to build.

So, let's ask it to build it out in Next.js. Please build out this website in a new folder in Next.js and Tailwind CSS. Make it clean with good animations. All models have now made a design of a website. So, I'm going to tell them to build it out. Please build out this website in Next.js and Tailwind CSS. That's all we need to give it. Transcribe that. Hit enter. And then I'm just going to copy this and paste it into Claude and Codex as well. And now we're going to have three different websites. And I'm very excited to see which one actually does the best job when it comes to front-end design, which is very important for me when I'm building our stuff. Very excited to see the results. There we go. All three models are now done. Let's start with Kimmi that's running in localhost:3000. And initial experience looks very, very clean. No intro animations, however, but if we scroll, then we have some animations. All of this looks extremely clean. I'm very impressed with the images that we have right here. This looks very cool as well, and you can see it's the sign so we can insert our own images. We have a testimonial section. We have our dentists. This is an extremely clean website that we could probably sell for anywhere around 500 to 1,000 bucks. It's also a high-converting website. It's not like we get overwhelmed when we go to it. Overall, a very, very professional website made by Kimmi in literally just a couple of minutes. And everything on the website is also working. We have clean animations, very, very complete from start to finish. Very impressed that we can build all of this, and we can build a bunch of websites like this for 15 bucks a month.

Let's look at the others. This is Claude Code. Also looks very good. Claude Code has actually gone in and found its own images, which looks nice. And then at the top we have this bar right here, which because I said it should be high-converting, is actually a good idea. This will convert more people, get more people to call. Also, right off the bat, a very good website for a local business. We have Google reviews right here. We have our services. Book free consultation takes into the form at the bottom, which also looks good. Let us scroll through. This looks good, too. Almost a bit too many different services, I'll say. Would almost overwhelm the customer. Then we have a dark mode section. This looks cool. This one right here, before and after, also looks cool. But of course, we would need to insert a good image right here. Then we have the people, the testimonial section. All of this looks very, very good. I almost like the other one better. We've been used to Claude Code absolutely dominating with front-end, but like this almost look better than this, in my opinion. It's more minimalistic. I actually like Kimmi's better. How interesting. Claude has already been best front end. And let's try Codex and see how that looks. Codex went with a bit of a different vibe. Still looks good. We have this little 4.9 out of five down here in the bottom right corner. Very impressed with all of these models, by the way, that they can create such a good website. But that's also because we gave it quite a good prompt. Then we have this card right here. This feeling, the feeling is mutual.

Also a very clean website. It's modern, it's good, it looks professional. However, some of the things are kind of off, like this text up here is too small. This number right here, like you wouldn't really get people to to uh to click this. Kind of a small website. I'll almost like it better if it was like a bit zoomed in. But of course it's not like that. It's almost like Codex in general has bit of a worse taste when it comes to design. And that's also why I've always used Claude code when it comes to the design before that. But like Kimmy is up there. I would say it's definitely just as good as Claude code when it comes to front end design. And that's absolutely insane. And I wanted to check my usage as well with Kimmy. And so far we've only used 5% of our 5-hour usage and 0.7% of our total usage. That is pretty insane. Kimmy is definitely up there, but let's talk about how I'm actually going to use it in my business and how you should use it, too. Should we cancel our Claude plans now? Well, let me tell you. Something very interesting is happening right now and all of a sudden I understand why after testing Kimmy myself in this video. This graph shows the Chinese AI models and how much they're used by US companies. And what you'll be able to see is this is skyrocketing right now. US companies are switching from the US model providers to the Chinese model providers, like Kimmy, like Deep Seek, like GLM. And why is that? Well, the main reason is because it's much, much cheaper. If we search up Fable 5 pricing and go to Claude's website, you can see that it's $10 per million tokens input and $50 per million tokens output.

And if you look at the Kimmy K4 pricing, it's $3 per million tokens input and $15 per million tokens output. So it is much, much cheaper, 1/3 of the price. So, it makes really good sense why people are switching when it's even less than 1/3 of the cost. But, what does this mean for us? Does this mean that I should go in and cancel my $200 a month Claude plan? It depends. Let me explain. Fable 5, which I still think is the best model right now, will only be available if you're paying $100 a month for Claude. And if you're using it via the API, then it's even more expensive. So, if you're like the regular AI user that spends $20 a month on a subscription, I would honestly recommend Kimiko over Claude. Because with a $20 a month Claude subscription, you will not have Fable 5, and you run into usage limits very, very quickly. However, I'm on the basic Kimiko plan right now, and my usage is looking really, really good. Usually, I have recommended CodeX if you only wanted to pay $20 a month, but honestly, Kimiko 3 is kind of just better. We proved that with both the video game and the front-end design. And if you told me that an open-source model like Kimiko would have been so good just like a month ago, I wouldn't have believed you. And this is both incredibly frightening, but also pretty good for us consumers. Let me explain why. Unless you've been living under a rock, then you know that companies like OpenAI and Anthropic has been raising an insane amount of money. Like, in March this year, OpenAI raised $122 billion. Surprise, these companies are not making any profit. They are investing heavily into the future of AI, which means they have been depending on being the frontier models, being the best models in the game.

Can you guess what will happen when a new model like Kimiko then comes out? It's open-source, which means that everyone has access to it, everyone can see what's going on behind the scenes, and it's 1/3 of the cost. So, I don't know if that's not very good for the companies that are investing heavily into their closed frontier models that they have spent billions and billions and billions of dollars developing. And don't get me wrong, it's good for us consumers. Competition makes even better models, and it brings down the price so we can have affordable AI. And that's also how I'll be using these models in my business. If there's an AI model that's just as good but 1/3 of the cost, I don't really care where it's coming from. I'll be using it. And I think you should, too. And on top of that, it's open-source. You could, if you had a million-dollar computer, run Kimiko free locally. Of course, that's probably not realistic, but you can still go and see the code of what's actually in that model, which is also pretty exciting. What's really important is that you stay up-to-date with all of these new AI things that is happening in the world. If you want to stay updated, then you should check out our AI community with 165,000 members. In here, you stay updated on AI, and we have full courses taking you through how to build a business with AI. I'll leave a link right below in the description. Thank you guys so much for watching. If you watched till the end, then remember to subscribe, and as always, have a wonderful rest of the day.