Transcription
Hello everyone. My name is Kostya. A great event has happened to us. 3.5+ has been released. This is a neural network with 1 million context, completely free, which we can use. Also, we have, of course, a paid API, for which for 40 cents we can, uh, buy 1 million input tokens and for 2 dollars 40 cents 1 million output tokens. That is, it is, uh, a little bit more expensive than GLM. Well, API is, of course, good, but we are interested in everything free in this case. Namely, what do we have? We have the Qwen chat. Where we can communicate completely for free with Qwen as much as we want. It understands voice input in Russian. It accepts a lot of things as input, pictures, JSONs, can conduct deep research, create images, create videos. That is, it is capable of a lot. Test it. But we are primarily interested in code creation, project creation using Qwen code. And let's look at the benchmarks. What do the benchmarks show? That it, first of all, has grown very strongly from its previous model, which is the green color. That is, in some parameters. And here are indicated as gray and light gray and even lighter gray. These are GPT52 Clot OPUS 4.5 and Gemini 3 Pro. That is, it has now become, well, practically on par with them, or slightly behind, somewhere in some tests it surpasses them. We are, of course, primarily interested in SVE. This is this one. And here it shows itself quite well. Well, maybe it's a little worse than them, but knowing Qwen, I have done quite a few projects with it in my time. Even on Qwen 3 coder. This is a powerhouse, that it has grown so much from its previous indicators. I think that in the case of Qwen, these graphs are unlikely to be drawn. It is meant that they are drawn simply from imagination. Most likely, these are indeed real numbers. We will be testing today without any APIs, without any kilocodes, completely in the trash. All these kilocodes we will test on native Qwen code. That is, this is a CLI that comes with Qwen. It is completely free to install. It is available in Russia. Without any hassle, without VPNs, everything as you like it. And what will we need? We will need to go through the installation, that is, install this CLI for ourselves. The following options are available via npm. Well, if you have Linux MacOS, then via QL. Well, I think you will figure it out here. Use it, if you don't know, just use npm and beforehand Node.js, of course, install it, as recommended here. That is, go to the link, install Node.js for yourself, then type in the terminal npm install. This one. And I, of course, already have everything, I will not type anything. And I will just open Qwen now, show how it looks. Well, the CLI itself, I mean. It looks like this. But if for some reason you don't like CLIs at all, if the sight of this terminal CLI spoils your appetite, then I have an option for you too. You can download Visual Studio Code. In it, we have this companion. You need to type Qwen code Companion in the search here, install it, and it will not differ from the usual one. For example, what you use code, raw code, cloud code, which is visual like this. That is, you also need to log in to your account here through login, and everything will work the same for you. All these commands will also work. You can also attach files here. Everything works approximately the same here, and modes, and so on. And I will return to the CLI. Let's go through the CLI a bit, and then I will tell you what tests we will have and what exactly we will be testing. So, look, this is a terminal program. I will make it bigger. What do we have available here? First of all, Russian language is available. That is, in leng, if we type, first of all, a slash, and here we have some menus to choose from. Here we can select settings, which model. That is, the model, look, we are immediately interested, well, what model do we have. And the model we have installed here, first of all, is free. This is Qwen+. This is very cool. I will choose it. When you open it for the first time, it will ask you to authorize. You will click AF and here you will select AOF. It will redirect you to the browser, you will authorize, and then the limits will become available to you. What else do we have here? We have agents, we have, uh, the same MCP, memory, skills, that is, everything is the same as everywhere, just like in Cloud Code. This is, you can say, an analogue of Cloud Code. Now I will tell you about the projects that we will now do, test. First, a game, a copy of a legendary game. This is Flappy Bird. I think you know what kind of game this is, but if not, then look. This is a landing page. Yes, yes, don't be surprised. This is, in general, a landing page for one of my products that actually exists. The only thing I changed is the name. Well, it's okay. There is also a mobile application. This is not some simple application, this is an application with an idea. Conditionally, there are movies, a certain list is known in advance, and you will have to present it in a certain way in the form of an application. So, we also have a web project. This is not an online store, not some nonsense. This is a single page application, that is, a single-page one, which performs one single function, that is, something like SaaS. The idea is that you upload a broken JSON there, and it gives you a not broken one on output. I have described this application very superficially. I think we will look at it later when it makes it. So, let's start with, I think, uh, with the game. So, I'm leaving here. Now I need to type SD so that the terminal understands which folder to go into. That is, I will drag game here. And here I will run Qwen again so that it understands what folder this is. Next, now I will open it for him. I have Project Game here, that is, a full description of the game, what is required, what is not required, which libraries. In general, everything is written in advance. I will now write create game from and the file. I am also attaching it, and that's it. Then it will start creating this game. That is, it understood that it will create a browser game Flappy Bird according to the specification. It understood that this is a spec. Everything is great. As soon as it finishes, we will, of course, look. And while it is doing it. So, it coped. It was done very quickly. In general, Qwen is characterized by the fact that it works very quickly and all the Chinese. So, here is the game. We open it, look, does it look like Flappy Bird. Very much so. Everything works, as you can see. That is, the clone, it has successfully made it, successfully implemented it. So, I try it. There is a score, there is a record. There is a restart button. In general, everything works, everything is great. It coped, it even did everything for us. We even have a bird. In general, this is the project we got. From now on, I will put all projects on GitHub, if anything. That is, you can go to my GitHub at any time, download this project, be sure to give it a star, and, in general, download it yourself, test it, I don't know, try to play it yourself if you want. The same will happen with the rest of the projects that we will be testing here. So, it's time for the next project. This will be a landing page now. Uh, so I'm closing all this. I type CD again. I need the landing page. Okay, we are in the landing page. And it's approximately the same. Create project from and I will specify project land. In general, everything is the same here. The specifications are all indicated, what needs to be done, what needs to be created, and so on. This is quite a large landing page, by the way. Let's see how it handles it. So, why not? I forgot to specify Qwen. So, open again. And now here I type create project and specify. Only from I will write. Ah, well, it will do. And so I press. And now it will create the landing page project here. As soon as it finishes, we will look again. Everything is the same as before. So, the project is created. It took 11.5 minutes to create this project. Well, the project is not small. Let's see what we got. So, this is the project we got. MVP in 14 days without chaos and unnecessary costs. Get a launch plan. So, full MVP preparation cycle. Idea development, audience analysis. So, okay. Something went very wrong here. Uh-huh. And here something has already gone very wrong. All the animation is, in principle, there. We need to ask him to fix it, because this is not acceptable. So, I wrote him a prompt about blocks in the graph. Well, I listed the blocks where nothing was displayed, and asked him to check the scrolling animation display and fix it. I'm sending it. Let him fix it. So, he fixed everything. We can look further. That is, everything is displayed for us. He showed it, right? Of course, the only thing is that everything works now, that's all, but as you can see, Qwen is not much of a designer. That is, in design, I see, he is not very strong, but he should perform well in code. That is, for design itself, it is better to use the undisputed leader Gemini 3 FL, as well as Mika 2 with po. That is, those neural networks, they really make cool designs. But this one made a good structure. Like, everything works not the first time, but it works correctly anyway. I know that you don't really like landing pages, so we will proceed to create something else. So, my terminal. So, we don't need all this now. I type CD here. to open the path. So, mobile or web? Mobile or web? Let's go with web. So, here it's approximately the same, only we initially have the entire structure, that is, it's a Next project, so that it doesn't waste time creating this Next project. And I will type N here first, and then I will type. create project from and here I will show the file. That is, here I have written completely what this should be, what kind of application, what it does, what can be used, and so on. That is, mainly this application is for logic, to check how Qwen writes logic. So, I'm sending it, let it do it. So, it understood, right? This is Jon Image Cleaner, we call this kind of app. As soon as it finishes, we will look again. So, it created the project for us. It spent about 12-13 minutes on it, 12 something it was. And we need to see what it created for us. To do this, we need to start the server. I have already started it in a separate window. That is, I just selected the CD folder and typed there, just as it asked. NPM run. Now we need to go to localhost. Like this. And we will see this Jon Image Cleaner. We have Drag and Drop, that is, we can just drag our Jon here, which needs to be cleaned. We have this test files Gon for this. We put it like this. Uh-huh. Yes, I see that it understands the essence. It performed validation. It says that the ones that are not broken at all. But partially not broken, we have 62, completely broken, we have 113 without images, and it did not count the completely valid ones. How many completely valid ones do we have? Well, it found them, but did not count them. I don't know why. Apparently, these values are not working well, but that's okay. Because it has mostly completely coped with what we needed. Aha. That is, I can select only valid ones. I can select valid ones, where at least one is valid. Well, that is, if there is at least one image, it is considered valid according to our technical specification. Uh, that is, you can select those where at least one image is valid, you can not select any, you can select only those that, well, I think you understood. Now we need to check how it performs the creation of a new JSON. So, image selection. What, what? Why is that? Where is select? Select invalid. And why is the button not working? Hmm, let it scan again. So, okay, it scanned again. I select valid or invalid, but the button for some reason, uh, load selected JSON. That is, selected JSON. And this is, in fact, our main functionality. It validates everything correctly, but it doesn't assemble what we need in the end. We need to write to it about this. In general, I wrote it this. That is, fix the bug, that this button is not working. I'm sending it. As soon as it fixes it, we will try again. So, it says that it fixed everything. We reload. We are asked to upload Jon again. We upload. Uh, any valid. So. Aha, everything, it validated. We scroll down, select only completely valid ones. Now the button is active. We selected 53 items and see what we got. Yes, a new JSON file has been formed. Now we can, you know, how to check this? We just reload and upload the new file here and see what's in it. And it only contains completely valid ones, which means our simple application that checks images works. That is, the application checks broken images, that is, broken links. If the links are not broken, it forms a new JSON. This is such a simple application called Jon Image Cleaner. Well, and the final test is a mobile application. Uh, so here it is, Mobile. We also have a project description here. Uh, so now we need to open the terminal, also type CD, select our path. to this mobile. Enter. Okay. And now our classic Create Project. Create Project. And from which file? This one, Project Mobile. So, only we forgot, only I forget to write. So, read. Okay, now I'll explain the essence. In general, this is something like a movie selection. Application. We have movies already in the assets, in this JSON file. By the way, the idea for this Jon Image Cleaner came about when creating this Movies JSON. That is, the neural network generated these files for me, well, JSON items, and not everywhere there were valid images. That is, it sometimes took from some places that the images, well, were not loading, they were simply no longer in the same places as it thought. And because of this, I got the idea for this web application. In general, now, based on this Movie JSON, it will create an application for us. This is where the movies are located, and it should, as it were, select a movie for us based on mood or completely randomly with swipes. As soon as it finishes, we will see what it has achieved. So, it created a mobile application. It took 20 minutes. Also, at the end, it asked that to launch, you need to type flutter run. I also pressed, typed CD mobile here, that is, I put it in and just typed flutter run. So, let's open the simulator. This is the mobile application we got. That is, on the first screen, by swiping, it offers us, I don't know, like in dating apps, to choose something. That is, I can swipe here, and it will offer us another movie. Forrest Gump, Pacific Rim, by the way, the photo doesn't fit. The Irishman, The Dark Knight, Black F2. That is, there is the year, how long the movie is, uh, what genre, that is, comedy, crime, comedy, crime, and a description. You can also go into the card. There will be, well, more details, the origin story of one of the most sinister villains, and so on. Who are the actors, who is the director, what is the mood. Yes, and regarding favorites, by the way. Let's try. So, favorites. I don't see it working. Ah, yes. But we also have moods. So, there are some layout issues here, but it's still okay. In general, we can select the mood here, for example, mysterious, right, and click, but I don't see the buttons. In theory, it should give us an application here. So, favorites are empty. And we need to write to him that based on mood, first of all, so that he fixes all these overflows. Secondly, our favorites are not working. Profile. What do we have here? Theme. Only dark theme. Language only Russian. Storage, export and version license. Well, it made a cool application anyway. The most important screen was, of course, the swipe. that we swipe, but it offers us some movie. That is, you swipe, a movie is offered. This is the most important thing that should work. By the way, there is also search, strangely enough. Ah, there is also search. Ah, and here you can even choose by genre, right? Cool. For example, military. And you can scroll. Aha. So, it all works like this. We need him to fix the movie selection by mood, and then everything will be great. In general, I wrote him a prompt to fix this screen. I explained to him the essence of this screen. Also, that there are overflows here, because the essence of this screen is that we should choose our mood, and it should select a movie for us based on that mood. This is how this screen should work. I hope he fixes it now, and we will look at the final result. So, he wrote that he fixed everything. Now I need to restart this application. Here, in the terminal, I just need to press Shift R, and my application will restart. So, I'm looking now, yes, now we have our mood, for example, I don't know, cheerful. and click select movie. And it selected the movie Han Solo: A Star Wars Story for us. We can click another option, and it will select another one for us. Well, that is, it only has two movies for this mood. We can choose something rebellious. This will be a different movie. That is, Deadpool fits perfectly, by the way. Ah, and that's it, there's nothing else. In general, this is how it should have worked. And now it works. So, let's move on to some final conclusions. Here we have Qwen 3.5+. As we found out, its context is 1 million, which is very good. We also found out that using Qwen code, we can use it absolutely for free. And there are huge limits, that is, I have never hit them. This is exactly what is written here about the limits. That is, there are 1,000 free requests per day. This is a lot. It will be enough for many things. The question remains the following. Can this be used as the main neural network for development? I think so. But with some reservations, that you will use at least on a minimal tariff some other cool neural network like Codex 53, GPT, I mean, uh, Cloud Code the same 4.6. Well, probably only these two. so that in case Qwen makes a serious mistake somewhere in some point and cannot resolve the problem itself, then you can resort to another neural network. Also, the architecture, for example, it is better to initially compose it on a stronger neural network, stronger than Qwen 3.5+. But in general, the neural network is very cool, it works quite fast. It copes with its tasks quite well. That is, I gave it three, well, four tasks, and it completed them all. Not the first time, but that's okay. As if from a free neural network, as if, as if, what did you expect for this money? As if these are still very cool results. I am very pleased. I will use Qwen every day now. I thank you for watching. I hope you liked the video. Be sure to give it a like. I tried very hard. Uh, subscribe to the channel, subscribe to the Telegram channel. See you soon. Bye. And be sure to write in the comments what you think about Qwen 3.5+.