NEW Qwen 3 Update (FREE!) 🤯

Alibaba have just released Quen 3. This is the latest model that just came out from Alibaba's company. And I've put together a comprehensive walkthrough to help you understand what this new model offers and everything you need to know about it. And what you'll learn in this tutorial is the key differences between Quen 3 and other leading models like GPT and Deep Seek. You'll learn how to access Quen 3 for free through their official interface and also how to code with it for free with their API. You're going to get a practical comparison of the three available Quen 3 sizes which are 32B, 30B, and 23. 5B. You'll get step-by-step instructions for running Quen locally using O Lama or LM Studio. You're going to get real world testing of Quen's coding capabilities, image generation, and video creation versus other tools like for example Gemini 2. 5 Pro and Claude. Additionally, you'll learn how to integrate Quen 3 with Visual Studio Code and Open Router. And on top of that, I'll walk you through hands-on examples showing you the model strengths and limitations. This is from

Alibaba, so it's a Chinese model, and it is free to use as well. So, you don't need to pay for Quen 3. You can see some of the benchmarks here and comparisons. So, you can see how it's outperforming versus Open A1, DC1, Grog 3 beta, 2. 5 Pro, and also Open AIO3 Mini. All right. So if we compare the benchmarks here, we can see for example Arena Hard and you can see that Open Air1 doesn't really compare to either of the new Quen models. And there are two models right here. So you've got Quen 3 235B 822B, right? Then 32B. I just feel like I'm reading out numbers here, but essentially what you can see is exactly how these benchmarks compare, how the models perform, etc. and I'll be showing you everything you need to know about these models. Okay, so if you want to get

access to this straight away, you can get access at chat. quen. ai and you can switch between the two different models. You've got Quen 30B and you've also got Quen 235B over here. So, as we get into this, let me show you exactly what we're going

to do. I'm going to run you through some of the key benchmarks. We got this presentation prepared for you. Preparation is key. And what we've got so far is the new Alibaba generation AI model family. All right. Now, the thing to note here is like these are hybrid thinking models and you can also connect MCPS which is super interesting. So, if we go inside the chat here, you'll see this new MCP feature that I believe is coming soon and that'll be released as well which is going to be super powerful. You can also for example code with this using O Lama, right? So, if you go to O Lama, you can host Quen 3 locally. I haven't tested that out yet. Last time I tested this sort of stuff out, my laptop literally couldn't handle it. But you can see all the different models and how you can host them here. So, you could probably host like quite a small model if you wanted to and then just start using that for building, right? Pretty amazing stuff. So, you can get access to this for free, brand new model. And additionally, you can get this on open router. So, if we go to open routter over here, open router. ai, you can start coding with Quen 3 inside the model section. If we go to the latest models, you actually got three versions of Quen 3. So you got 8B, which is free, 14B, which is also free, and also Quinn 32B as well, free. So you can actually code with the API. I'll show you how to do that later on. Also, we'll compare like some of the models to see how they perform together. Just as an overview here, Quen have been on my radar for quite a while. I've been looking at their models since they released Quen 2 really. That's the first one that caught my attention and ever since then like the models are really good. The cool thing about this is you get a lot more functionality inside the chat with Quen, right? So if you go into the chat with Quinn here, you can actually for example upload audios, videos, etc. You can use that artifacts feature. They also have a deep research feature that's coming soon as well as image generation and video generation which is quite different. Chat cheap is not going to generate videos for you directly inside the chat. Additionally, you've got some other stuff like you can check the news. So, you can actually link this directly to search which means it can search the web and that sort of thing which is really cool. You can analyze images, summarize text, etc. But the main thing to note here is like you're getting a very powerful model. You're not paying for it. It's very similar in terms of benchmarks to something like Deep Seek and actually outperforms Open Air01. And

also, this is a hybrid reasoning model. So, it can decide okay when to think or not. And also, it's connected to the internet. So, it's a very powerful model, very underrated, and like 99% of people out there never going to know about this model. But if you're watching this video, you understand it already more than most people, which is great. So, let's get straight into these models right here. So, Quen represents Alibaba's most advanced AI model family, featuring both dense a mixture of experts architectures under an Apache 2. 0 license. All right, the key highlights here basically number one, it's trained on roughly 36 million tokens, nearly double the data set size of Quen 2. 5. Number two, it supports over 119 languages and dialects with extended context lengths up to 128K tokens. It's got a unique hybrid thinking and it's able to switch between reason. So, we're seeing this across like chat GBT also claude and that sort of things like you can switch between these models. You don't have to pick one or the other. You just pick which one is the best for you. And also you can typically like understand okay which one is the best based on what you're searching and how you're using the models. Okay. Next up improved agentic capabilities and tool calling functionality. So you can see like the performance breakthrough here in terms of what's been achieved and also you've got the summary in terms of the context. There's many different models. You got large, medium and small. Okay. So obviously the smaller model has a lower context, the larger one has a higher context and the larger one is more powerful. Now in terms of hybrid thinking modes, so Quinn introduces a new sort of dual mode operation that allows the model to either think step by step or provide immediate answers based on task complexity. So it's kind of like a seamless mode where it can switch between one or the other. And it can choose between thinking mode or fast mode, right? So fast mode obviously is like non-thinking mode. It's kind of like the human brain, right? I don't know if you ever read Thinking Fast, Thinking Slow. Great book, by the way. Super long, but great book. And basically, it talks about how you've got the impulsive side of your mind, like the lizard brain, and that could just answer stuff really quickly. Then you've got like the thinking and reasoning mode. And basically, what that does is when you sit down to think and you reflect and you spend a bit more time, you don't just make some crazy impulse decision. That's basically what the AI can do as well, right? You're switching between two different modes, very similar to the human brain in many ways. And then you can see here's a thinking example. So it will do like a think mode. And then if it's a more complex task, that's what it's going to do. And then in terms of the thinking budget, performance improves, right? Performance versus computation. Now let's have a look at benchmark performance here. So you can see in terms of outperforming other models on the market right now, you can actually see down here. So 03 Mini is being outperformed by Quen 3 by 3. 2%. 2% Google 2. 7% Deepc Car 1 1. 8%. So it's marginally better. It's not like destroying these other models, but it is super powerful. And bear in mind again, this is free. I think they're planning to charge eventually when they've run through the testing periods and the preview periods on their models. 03 Mini obviously you have to pay for that and it's a paid model and you get limited like you're never going to be able to use API for free, but you can use the API for free using Quen, right? is also outperforming DeepSeek R1 terms of mathematical reasoning. So you can see here Quen 332B versus 01 is up 5. 4%. Quen 38B versus Quen 2. 5 is up 2. 1% and Quinn is outperforming 2. 5 as well. And then the live code bench results over here you've got so versus deep sky 1 on the live code bench results scores 76. 9% and Quen 3 scores 78. 4%. What does that mean? That means coin is outperforming it by a couple of percent, right? Not a massive amount, but you can see the difference here and the progress right here as well. Now, I was talking about

language functionality. So, you've got a breakdown of all the languages that you can use inside Quen right here. But it's got many different languages. So, for example, if you're coding in Arabic, you can do that. You've even got like Dravidian and Turkish languages, additional families. So, for example, like Japanese, Korean, Thai, etc. Even Laos and Haitian. And then you can code in Chinese, simplified, traditional, Cantonese, etc. Right? So basically all the main languages are covered inside this model as well. So if you're trying to code in multiple languages, you can. And that's really powerful. There's a lot of SEOs who watch this channel and a lot of those SEOs use foreign SEO, right? And you can build websites. So you can build projects or apps not just in English, but in multiple different languages using this new model as well. And then now let's talk about the agentic capabilities. All right. So popular tools, you got the web search, you've got, for example, image generation, browser use. I haven't tested that out, but it does say that you can use browser use. I'm not sure about that, but I would just be I would test this. Some of this stuff has not been like fully released, right? And what I mean by that, for example, is like inside Quen, there's a coming soon option for MCPs inside the chat, but you can't use it yet, right? So like some of the features might not be fully functional yet. So you can see how this performs stems of that and then also we've got the pre-training over here as well. Right? So this was trained on nearly 36 trillion tokens approximately double the data set of Quen 2. 5. And then you can see the model architecture right here. It's actually it's interesting to see like the training pipeline for these models as well. So you got basic then you've got the knowledge enhancement then you got the context extension as well. Now in terms of deployment right like how can you use this? You can use this inside cloud services. So for example like Alibaba cloud model and hugging face interface API you can use cloud as well right self-hosted. So you can actually use this for example on oama you can download the model locally and do that directly and also for local development. So yeah again you can use llama and additionally you can use this inside the chat as well. So you can go to for example chat. quen. ai and then start using this. By the way, if you want all the video notes from today, I'll put this inside the AI profit boardroom inside our school community. So if you want all the notes and everything else, then you can get that inside the SAP section here. I'll just add a new page for Quinn three. Let's pop that in. Which you can get directly right there. Okay. It's got a bit of information about finetuning that sort of thing. If you want to fine-tune the model, right, that's really good for like legal document analysis, customer support, so you can train it on all like your custom knowledge and all the SAPs, all the systems you have inside your business, all the training, all the FAQs, etc. Like for us, for example, as a business, we have a lot of documentation on FAQs, how to answer certain questions, how to run sales calls, client calls, etc. We have playbook for pretty much every department, right? And so when you fine-tune your own models, you can actually use this really well for customer support. Codebased specific assistance that could be really cool. Healthcare specialization is super interesting as one as well. So for example, like people can fine-tune the AI model on medical literature and clinical notes to create a healthcare focus assistant with like domain knowledge, right? It's going to be like a specialist essentially. And then in terms of practical applications, how would you actually use Quen 3, right? So there's all this stuff you can use it for like for example creating software, customer service, legal, education, finance. You can also automate a lot of stuff. So you know you could build this into web applications, mobile apps, chat bots, etc. And in terms of the future road map, right? So if we're looking at like long-term, right? So, I just I saw that someone said, "Call me Troy said, "You guys realize you're talking to an AI avatar. " You are not. Don't worry. This is a real Julian Goldie in the flesh. So, as we keep going right now, you can see some ideas in terms of how it uses, what it's going to be used for in the future as well. So, you can imagine like this is getting closer and closer to AGI. I also think like with the MCP models that going to be implemented into this as well, you're basically getting like an autonomous agent system, especially if for example those MCPs linked to browser use and computer use, then it can just go off and use your browser without you. And then you can also create like these sort of specialized domain experts based on the fine-tuned models. Okay, so let's

get into starting to use it. If you want to use this inside the chat, you can just go inside chat. co. ai AI and then you can select between three different models of Quen 3. So you've got Quen 3 32B that is the most powerful and the most dense model right then you've got Quen 330B A3B that is a complex compact and high performance mixture of experts model and then finally you have Quen 3235B A22B right the most powerful mixture of experts language model. So you got a few choices there in terms of what you want to use how you want to use it etc. And we can just give this a cheeky world to test out, right? What we can also see inside here is you can flick between thinking and you can also control the thinking budget, right? Which is something I've never seen introduced into an AI model directly. I've seen it obviously inside APIs and that sort of thing when you're coding, but I've never seen the ability to be able to select your thinking budget. It's pretty amazing. So, what this means is like you can choose how many tokens you assign to the query that you're running, right? Basically, if you wanted to think more, then you increase the tokens. You wanted to think less, reduce the tokens. If you want faster responses, you would reduce the tokens. You want slower responses but better thought out, then you would increase the tokens, right? And so, for example, like we can just keep that pretty low at the minute for now. You can click that on or click it off as well. And then you can also upload documents, images, videos, and audio, etc. as well. And then inside here, for example, let's just take a prompt from the master prompts inside the AR profit boardroom just to test this out. So what we're going to say here is make me a captivating endless runner game. Key instructions on the screen, blah blah, right? And basically this should build out a sort of endless runner dinosaur game, right? So it's coding out. The speed at which it runs is pretty awesome. Again, we haven't really selected thinking mode here, so we're not running the full power of it. If the outputs are not that great, I wouldn't be surprised. We'll check it out in a second. I'm going to run this on p5 js and see how it performs. You see how fast it is, right? It's absolutely rapid. Bear in mind as well, like you can generate videos in here. I don't know if you're not using YouTube in 2025 to grow your business, I would say you're crazy cuz it's just such a good opportunity. And we have a lot of AI related videos or AI automated videos on YouTube that actually perform very well, sometimes outperform our human capabilities. So the cool thing about this is you can actually just create videos inside Quen. You could stitch them together if they're too short, but essentially you should be looking at different ways you can leverage Quen to get the best results. So if we paste this code in right here, then we're going to play it on the preview. So what I'm going to do here, I'm going to run exactly the same prompt right inside Quen 32B. We'll start a new chat and then we're going to switch the thinking budget up here as high as it possibly can be. All right, so we'll paste that in and I just wanted to show you like the difference between thinking and non-thinking mode. So we'll see that in a second. Alex Wise asks, "Hi, best model for coding so far. " So, for me personally, from what everything I've tested, and it always varies because, you know, when these models get released, they're really good and then they just drop down to base model, if that makes sense. I don't know if they get tuned down a little bit when it gets a bit too expensive. But from what I've seen so far, I would recommend using Gemini 2. 5 Pro. It seems like the best model for sure, and it's also free to use. You can go to a studio. com, and then you can code directly in here for free using Gemini 2. 5 Pro. It's also pretty good inside Gemini Advanced as well, but if you want to use it for free, just go inside Gemini. I also like Claude as well, but if you're coding on the API, it gets super expensive. And also, I just tend to find better and faster outputs from Gemini 2. 5 Pro. So, you can see you're coding out right now. Again, it's going to take a lot longer, right? Because we're using a higher budget for thinking tokens. So if you looked previously at the example right there, this took really this was really fast to reply the endless runner game, but it doesn't really work so well, right? Whereas for example, we can run the thinking budget here, increase the tokens. So I wouldn't recommend using full token limit every time and then you're going to get a much better much more thought out response. So wait for that to code out and see what we get back in a second. You've also got here a artifacts feature. Now this is similar to for example claw artifacts or for example chat GPT canvas Gemini canvas where you can preview the code inside the chat if you're using something like HTML right so if you're running HTML and you're coding in HTML then you can preview and generate some tools with that in a second I'll show you exactly what I mean using some of the master prompts from the AI profit boardroom in a second. Yashaz is asking how is cursor sir. I'm not a big fan of cursor to be honest with you. I just Yeah, I'm not gonna say much more on that. I wouldn't for me personally out of everything that I've done. You probably saw me building an image app the other day which GPT image one. If you haven't, check out the video on YouTube. But essentially, if you look at the way that I coded that, the only time it worked was inside Visual Studio Code. When I've tested it on Bolt, Lovable, Windsurf, when I've tested trying to build the same thing inside, for example, other tools, it just doesn't work as well. So, I'd recommend Visual Studio Code. It's the best. It's free. You can plug in your API inside there, and it's super easy. So, I'm going to wait for that to load right now. We'll see what we get back here. Basecam Bernie says, "I am running Quen 3 32B on my two GPUs, Olama, and this is a clever little model. " Nice. Yeah, I'm always loving Quen's updates. I think they usually create really awesome stuff. Let's just make sure this actually works. So, we're going to grab this code right here and just test it out. Now, bear in mind, like this has only come out a few hours ago. I'm testing it for the first time, right? So, if it doesn't work perfectly, we'll check that together. It seems to work pretty nicely so far. So, we've got the output of the game right here. This is the code. We're running this at editor. p5js. org if you want to check it out. And then we've got the game ready to go. So, let's see how this goes. Yeah, it's not bad at all. I don't like the UI so much. I always think like Gemini 2. 5 Pro and especially Claude are best for coding out the UI of stuff. Quite often what I'll do inside Visual Studio Code is like code the back end with Gemini 2. 5 Pro and then code the front end with 3. 7 Sonic, but it's not bad at all. The game is fully functional. It works. It's got an interesting UI and there's no bugs on that game at all. Pretty cool. Super easy. Queen Harper says, "How much RAM is recommended to run the game locally? " I think it depends which model you download and use. Right. So if you want to host stuff locally, if you want to run it locally, what you can do is you can go to olama. com. You download oama like so. It's completely free. It's available for Mac, OS, Linux, and Windows. Then from here, just go to the model section and if you want to run this locally, just click on Quent 3. Now, why would you want to run something locally? Let me explain that to you. So number one, you can actually code if you run something locally, right? So if you build something locally, you can code locally using O Lama as well. You just plug that in the API settings. Number two, obviously it works offline and you could be flying on a plane. You pull out your laptop, you run O Lama and the Quen model and then you can use it. Super useful if you're offline. And yeah, it works completely offline as well. And then also it's private, right? So everything that you create with that local model isn't fed back into the cloud or anything like that. It's all confidential. So, let's keep going right now. So, if we have a look down here, let's see what we got on the information. Doesn't really have much details on like how much you need to run it. I'm doing a live stream, right? So, I've got quite a lot. I've got two instances. Chrome running. I've got Streamlabs running. I got all sorts of stuff running at the minute. I'm pretty sure if I try and run one of these bigger models, my laptop is just going to overheat and go crazy and slow down. It's going to ruin the whole stream. But what we can do for example is we can select like 0. 6B because I think that'll be pretty chill. Then we can copy this model name over here. So you just click the dropown you can select which model you want, which size etc. And then inside this section you would just click copy. So on the Quen 3 model section right now once you've downloaded OAMA you can then open it up. Right. And once you've opened it up you will see this Olama icon over here. So you see that's llama running. Make sure you got it running. Then we can go into terminal. If I remember correctly, it's been about a month since I last did this, but we just paste that in there and that should download for us. Right. So you can see here, for example, inside terminal, that is now downloading the model. Now, if the stream starts slowing down or if this video of the laptop starts overheating or something like that, I'm just going to shut down the terminal. But that is pretty fast and easy to download. Really simple. I'm running this on a MacBook Pro M3 that is actually not loading. Look at that. Unable to load the model. All right, let's try the 0. 6B and we'll see if that works. It's good for you to see this stuff live. I always like to show my mistakes as well because if something doesn't work for me, then I want to show you that cuz that's as important as knowing what doesn't work, if that makes sense. So if you see a video from me and there's a mistake in there or something doesn't work. I think that's great because it's going to save you a lot of time. If you see what doesn't work, then you can avoid that. Use what actually does work. So let's see if we can run the 1. 4 GB model instead. Otherwise, what we might have to do is up is actually update OAM. But we'll test out in a second and see how it goes. So that is 90% away there. Should be good to go. Spilling it in. Is it going to work? Unable to load model. Interesting. I've never seen that before inside a llama actually to be fair. That is very interesting. Maybe it's Why is that not working? Let's see if we can try running it again. No. No bueno. Maybe we could run the 4B model, but that's not working for me. So, that's just something to be aware of. I've never actually seen that before. Like on the older models, usually you don't have any issues right there. So, it's quite interesting to see. And we'll keep going through now. All right. So, inside the Discord here, let's open this up. See what people are building with it as well. I'm just going to try and update Lama cuz I think that might be the problem. So, let me just go on to Lama down here. We'll download that. We'll try and run the latest model or we can restart update. Ah, that's why. So, we'll restart now. All right. Should be all good now, hopefully. Let's check this out again. Copy that. Plug that in. Hit run on that. Yeah, that's working now. All right. So, just something to be aware of there is if OAMA isn't working for you, what you want to do is click on the OAMA icon and then hit restart to update, right? And then once you run the model again, it'll start working for you. Shout out to Basec Camp Bernie for just giving the tip in the chat there on how to do that. Awesome. All right, so we've got that running. This is Olama 0. 6B 6B over here. Let's check this out. So, we'll just say, "Are you working? " And that's working. Beautiful. All right. So, let's see. You can see here it says, "Sure, I'm available to help with your questions or needs. How can I assist you today? " So, basically what we've done here is we've got Quen 3 running through Llama. We've got it running locally and it hasn't cost us a penny again, right? So, what we can do from here is like I say, okay, what model are you? just to double check. And there we go. It actually doesn't know its own identity, but that's pretty common with language models to be fair. And then you can just copy this. And anytime you want to run it, just copy that and you can run it. And you've got a list of all the different ways to run the models down here. Right. The good thing about 0. 6b is it's pretty lightweight and it's quick to install. The biggest problem with these models is that the more lightweight the model is, the less powerful it's going to be, right? So obviously if you were using something like something that's 142 GB like you see right here that's going to be way more powerful but for most people they can't run it on the laptop. It's just going to be too big and it's going to slow everything down. So I'm just going to quit the terminal there. That's exactly how to run this stuff locally. So we'll terminate that and then we're good to go on that and we're back in the game. All right. A lot of people saying LM Studio is better than Alama which I've not tested out. Let's have a look here. So this looks like another way you can run local AIS. To be fair, the UI is way better, isn't it? Way better. Cool. I'm going have to check that out at some point. Little tip for you there. So, go to lmstudio. ai and then download that. That seems to be nicer UI for sure. All right. So, if we go back here, I've shown you how to build and code inside the chat directly. You've got choices between three different models. We can check out the video generation if you want as well. So, we'll just put like cat eating biscatinos. It's your classic use case for using AI of course and we'll see how that goes. Usually that's going to be pretty slow. So you can see the speed at which that's running at but at the same time it's free, right? And you can always have that running in a separate tab and you don't have to use that directly. Okay. So we've covered how to run this locally. use it inside the chat. What's next? So

what we can also do is you can get the API key and then start coding with that. Let me show you how to do this. Right. So, for example, if we go over to let's have a look here. We'll open up Visual Studio Code. If you don't know how to use Visual Studio Code or client or anything like that, just download it. You can just type in on Google or just check out the SPS inside the A profit boardroom if you want to figure this stuff out. Got tons of tutorials on it. And then inside here, we can start changing our settings. Right. So, if we go to client over here, and then we're going to go to settings. And then we can actually select the API key. Right? So you can actually see we've got Alama running there. So you can plug in the settings for Alama if you want to. You've got LM Studio as well. So you can run LM Studio locally and then run the model there. You've also got Alibaba, so you can get an API key from Alibaba for Quen. You just plug in the API key here. That's another way. My favorite and easiest way I think honestly is just to go to open router and then plug in the API key here. and then select your model. Right. So, if we type in here, let's see if Quen is updated inside the chat. Have we got Quen 3? Yeah, there we have. And then you can see there's tons of different models here. It's absolutely ridiculous. Right. So, you've got like 30B, 23, 5B, you got 14B, etc. Right. So, for example, if we select Quinn/quen 3, 14b colon 3, we can select that. And then from here, we can start coding with this. Right? So if we hit save. So we'll just check are how you working. I just need to make sure I've got the right API key in here. Wait for that to load. Also the other thing you're going to note here is if you're using the thinking model sometimes going to be a bit slower. You know what I mean? If it's got a think on every single IP API request, but you can see here that it's done the thinking mode. So it's like here's the thinking model and then we've got the API request model back. So this is quite interesting because you can now switch between all these different models and then figure out okay which one is the best for coding that sort of thing. So let's test out an example prompt inside here. What can we say? I'm going to take this one right here. All right. So we're going to say generate a high converting landing page. Just build out like a onepage landing page inside client using quen 3. If you ever want to check like which API you're using from open router, click at the bottom of client, right? So at the bottom you can click client and it's got open router coin 3 and then you select whichever one you want and just make sure you're using the free models if you want to use those. Right? So from here let's test this out and paste this in. We'll see how that goes. I'm pretty sure as well like you could quite easily link this to MCPS inside Klein and then use that as like a more agentic framework, right? So if you want to make Quen way more powerful for coding, all you need to do is just set up some MCPs, connect the MCPS to the API and then you're going to get much better outputs when you're building stuff, right? So for example, like Brave Search or Plexity or if you're coding a lot, then Context 7 is pretty good. You know, from here it's pretty easy and simple to do. So you can see it's now coding out the project. And if you want to see, okay, what's it doing so far? So we've got this section here where it says okay here's what I'm going to do right so if we click on thinking we can see exactly what it plans to do step by step also root code is pretty good for that because root code it tends to give you like a project plan if that makes sense and then using boomerang task it will create a separate task for each part that you're coding out and also what I like to do is select auto approve here just so that it can run without me I don't need to click approve every single time it'll save you a lot of time when you're using out client right there right so what we've done right down is you can see here it says I've created a high converting landing page for Goldie agency with a vibrant hero section CTA blah blah I would just take this with a little pinch of salt right so the reason that I say that is like when you're coding out with this with Quen and you get the first draft from Klein usually the UI is trash right but the functionality works so let's run this and see what we get back when open up yes you see like the UI is very basic but it has created the landing page, right? So, we could do a lot better with the UI. What I typically like to do when it comes to coding out the UI instead of the actual content is then use something like Claude to

improve that. All right? So, I can use something like Quinn to just code the basics, etc. And then we can switch over to Claude. So, if we type in Claude, which is not free, by the way, but it is super powerful. So, if we type in Claude 3. 7 over here, and then we'll close that. And then in here we can say, okay, make the UI super modern, sleek, and beautiful. And what that'll do is just run through the code you've already got from Quen for free and just give it a nicer interface. And typically, like I say, like Ford is probably the best for coding the UI and stuff out for now. Interesting comment from Base Camp Bernie. He says, "Quen 3 is nicely optimized for MCP use, right? So it tends to pick tools. I haven't tested that out, but it' be quite interesting to see that in a second. We'll see what we get back from here. So, you can see now like Claude is just going through the files and just improving it, updating the CSS, etc., the HTML, just creating a better version of what we've already got. And honestly, I don't think you're ever going to get the same UI from Quen you are from other tools. So, that's one way to do it.

Another way to do it if you want to code for free with the API is you can use something like bolt local bolt DIY. Right? I'll show you how to use this in a second. So if we close this up, we're going to quit Visual Studio Code. I'm going to shut it down cuz it just takes up a lot of memory. We're going to go inside the A profit boardroom. We'll type in Bolt DIY. And then from here, one second. From here, we're going to go to the Bolt local SP. this inside the AR profit boardroom which you can see right here and if we open up the terminal now we can actually just run bolt DIY locally and then we should be able to code with quen free inside there now the reason

that I would use something like bolt local for coding instead of visual studio code is like again you saw the UI a second ago right it's not that great so what you can do instead is you can run bolt local and that tends to get better results. All right, so we'll run this now. We're just going to hit npm rundev on there. Pretty easy and simple to use. If you use a the SAP from the air profit boardroom, you can just literally just copy and paste these commands. Like you can see right here. I'm not a coder, but I know how to copy and paste, right? And that's all you need. So from here, what you can do is you can select between all these APIs. And the one that we're specifically going to use is open router. And then I'm just going to see if we can use Quen 3. Yeah, we can. So, it's already updated in there. So, you see here in the drop downs, you can switch between all these different models. It's got all the free models as well. So, for example, 14B, 23, 5B, 30B, etc. I think on the website it said 30B is the most powerful. So, we can select 30B right there. And then we just need to set our API key. All right. Okay. So if you click on get API key, we can get one from open router and then plug that into bolt local. Now the difference between using bolt DIY versus something like Bolt DIY allows you to plug in your own API key, right? Which means that you can use Quen 3 directly in there. So I'm just going to grab an API key in the background. I'll show you exactly what I mean. So, I've copied the API key and then I'm going to plug this in this section right here. We'll paste that in. Hit enter. There's also a bunch of settings you can change in here as well. So, if we go to bolt and then we click on the settings cog in the bottom left, you can export all your chats, that sort of thing. You can switch between providers and you can switch them off if you don't like the drop down if there's too many. You can also if you're running for example or LM studio with Quen then you can plug in the base URL over here. You can connect it to GitHub and there's also a bunch of features right here. So for example like I debug auto select code template right which lets Bolt select the best starter template for your project and then context optimization as well. All right so let's off that and from here we can start building stuff out. So, let's create something super basic for now. I'm just going to say, okay, create a retro keyboard. Again, we're using Quen 3 to build this out using Open Router, and it

is free. Bolt is free, and Quen 3 is free as well. So, let's see what we get back in a second here. Samuel King says 232B is the most powerful. All right, bear that in mind. I might use that in a second if we don't get good outputs from this. Kaisto says is Quen 2 free. Why would you use Quen 2 if you got Quen 3? That's what I don't understand. I would recommend using Quen 3 instead. You're going to get the best outputs there. So, we got a terminal error here. Let's just ask Bolt again. Get that fixed. Now, you can switch if one of these is not working so well. So, like you can see right here, then you can switch between API keys during the actual video as well. Right. So, we've got the model right here. This is the retro terminal. So, we can type with the keyboard as you can see. And that's working first time around. Again, I don't like the UI that we're getting from this. But let's go back to BoltDIY and we'll switch now to I think it was 232B is the most powerful model. But let's try 32B then for now. Oh, 235B we've got. Yeah, let's try that. All right. 235B. So, we'll try that for coding instead. And then we're going to say, okay, build out a I'm tempted to write a self-playing snake game, but I just don't want to be one of those cliche YouTubers. I think I've passed that point. So, let's see instead what we could build. We'll take this one right here. By the way, we got the video back from Quen 3. Now, let a cat eating Biscatinos as a video. Incredible. What a time to be alive. It's not bad to be fair. It looks a bit laggy, but I'm sure if you download that, the preview will be much better. So, let's go back to Bolt now and we can plug this in. We can also use the AI to improve the prompts and you can actually speak with your mic and then put in the inputs right there. So, let's hit enter on that and see what we get back. Animated intelligence said, "Just jumping on here. Is there a video on how to set up bolt? " Yeah, check out the AI profit boardroom and you'll see. Or if you just type in on YouTube Julian Goldiebolt DIY, you'll see tons of videos from me on how to set this stuff up, right? And exactly how it works. And regarding Lovable, I know you said you spent a lot of money on Lovable. I tested out Lovable 2. 0 on a live stream last week. Wasn't massively impressed. I'll be 100% honest with you. From what I've seen, the best output so far tend to be Visual Studio Code with claw 3. 7 Sonet. Bolt new is reliable. Occasionally, you get these errors that you just can't fix and you have to start a whole new chat. But bolt DIY is not bad. It's not bad at all. If you just want something basic, you can plug in your own API keys into there. Then you can start using this sort of stuff. So, you can see that's coding out now. We'll see what we get back in a second. I want

to see what people are saying about Quen 3 as well on Twitter. So, let's type in quen 3 over here and see what people are building and creating with it. This is one of the best ways to do research. So, you can just type in whatever you're trying to look for inside Twitter or x. com, whatever you want to call it. My brain still calls it Twitter even though it should be X. And then you can see what people are building with it and how it's working, etc. You see my live stream right there popping up and we'll see what people are building. So, if we switch to latest now, see if anyone's created anything impressive with it. And also it will show you like all the benchmarks and what people are saying about it. It's one of the best ways to do research. So let's go back to Bolt now. Oh, what's happening here? What's going on here? It's totally broken on me. It's a terminal broken. Maybe I just go back to the chat. So I'm just going to say inside here, keep going cuz seem to break, which is crazy, huh? Maybe we'll close some of the tabs. Let's see what else Quinn have to say about it. So these are some of the key benchmarks. Again, what's interesting as well is like you can see the difference in power versus something like non-thinking mode, right? So non-thinking mode, obviously it's a hybrid reasoning model. So non-thinking mode, it's like a baseline red at the bottom. And then thinking mode, you can see how it performs on like maths and that sort of thing as it goes up. Just wondering if it's slowing down on the API because it's too busy or something like that. Let me try the paid one instead. I think less people will be using that. So, I'm going to try the paid one here. Again, you can use this for free. I think it's just because it's been released. So, the API is taking longer. There we go. Look at that. Way faster. It's going to shut alarm down as well. Yeah, I think Bolt is just totally breaking on me. Bolt DIY. It's simply It's because I've got like the streaming software and everything else running in the background. So, you can see here it's super slow. But I don't think that's the thing that's actually bolt just breaking on me because I'm running it locally. Maybe I'll simplify the prompt and then use the API in a sec. It's broken. All right. My laptop's too slow. M3 with Streamlabs running. Not great, but you can run that locally on yourself and it will work a lot better. All right. Now

if you want to connect this with MCPS as well, we can go back to Visual Studio Code. Let me open this up. I'm going to stop the terminal now. Terminate that bad boy. And then inside terminal, if you go down to client, go to the MCP marketplace, check what's installed. So you can see here we've got Black Research and Context 7. We can also add more MCP servers. So for example, like browser tools. What else we got here? Browser use, Brave Search. So, let's install Brave search as well. MCP. So, that's how you can install MCPS. We'll see if Quinn can install MCPS for us. We'll hit run command on that. Basecam Bernie says, "I love MCP so much. I have a solid rotation of the favorites. " Let us know what your favorite MCPs. Would love to know. So, I'm going to click on provide API key. And then if you want to grab an API key from Brave Search, which is free, you can just go to Brave Search over here. Basically, what you're doing with Brave Search is like connecting it to the internet if that makes sense. So, we'll log in now. Grab that. Probably going to have to verify that as well. Yeah. So, we've got to verify that the API key code. There we go. And then you just go to API key section over here. Add an API key. You can put for example client. You can select your subscription. So, I've got free here. And this will just allow you to set up the API key from Brave Search. I'll delete this after, but I'll hit enter now for this API key to start using it. Now, you can see it's programming that in. Let's see if it can install it. To be fair, even on the paid API, I'm using the paid API today just so that it doesn't let me down whilst everyone else is using the free one. But you can see here the cost of it is super cheap. So, it's not even been like a dollar. It's 0. 00 16 to start using this so far. So, let's hit save now. And you can see here inside the MCP settings, we've got all the details of all the tools that we're using, right? So, if you ever want to edit the API keys, delete them out, etc., you can do that inside the MCP. Everyone's saying contact 7 is their favorite API key. Yeah, I get that. All right. So, let's hit save on that. Then if we go back to Visual Studio Code now, here we go. Boom. Brave search working nicely. All right, let's see. So we'll say, okay, use Brave Search MCP and research all the useful facts about Quen 3 released today. So, it seems really I mean like honestly when I've used like Gemini 2. 5 Pro to set up MCPS, they often fail me whereas this actually seems to work. So, we're going to approve that. Just going to make sure it actually works. In the

meantime, let's check out Quen's image generation. So, if we go back over here, go to Quen, then we'll put in image generation. And let's just test this out. So, I'm going to use this prompt for the Howling Wolf. see what we get back. The prompting is really interesting actually when you're using Gwen. You see how fast it is. That's way faster than chat GBT. And there we go. All right. So, pretty quick for image generation. Let's just check that versus chat GBT. So, I'm going to say inside chat GBT40 create image. Plug that in. We'll see what we get back and compare that side by side. Mate, so far Quen is not very good at setting up MCPS. It's actually terrible for setting up MCPs. I'm not going to lie. I don't want to be SE and tell you it's good. But look at all the back and forth here. This is worse than Gemini 2. 5 Pro. So, what I'm going to do now just to get this working, we'll go to Anthropic Claw 3. 7 Sonet. Then I'm going to say install the MCP Brave Search and check it's working, right? And then from there, we'll test if Quen can actually use it after. But it just seems like for installing MCPS, especially on client, brave search is absolutely terrible. Quen 3 is absolutely terrible. All right, so let's now compare the image generation of Quen versus Chat GPT. Which one do I like the most? I would probably say Chat GBT's image generation is a lot more advanced. That image looks so cool. Let's compare that. They're not bad, but I would genuinely say like the chaty image over here still looks much better. Much more interesting. It looks a lot more like stylish and interesting. Whereas this one, like you can see even the moon was cut off of there. It's got some weird like artifacts on its body which don't seem to match up properly. Whereas this one just looks a lot more lifelike and realistic. Edward White says, "Don't show API keys. " It's all right. I'm going to delete them after the stream. So, it's no big deal. Look at that. So, Claude works perfectly when you set up the MCPS whereas Quen 3 was just terrible. Let's check it out for actually using MCPS now. So, if we switch over to Open Router Quen 3 14B and let's see if we can use 30B instead just get better outputs or 32B. And then what we'll do here is we'll say the same thing. All right. So research she's in brave search the news about quen 3 today and that is working perfectly. All right. So it seems like for setting up the MCP I would definitely recommend using Claude. It would just save you a lot of time. And then for actually calling the MCP quen 3 seems to work really well. You can see all the details right here. It's linked to the different sources. We've got all the information. And now you've got the Quen 3 API plugged into Visual Studio Code. It's super cheap to use or you can use the free version and then you can call MCPS with it directly like you can see we've done with Brave Search. Just going to go back. If you want to learn how to delete your API key on Brave Search for example, let me show you how to do this. So we can go back here, log in. By the way, if the screen was just removed for a sec there, sorry about that, but it should be all good now. So let's get the verification code and then we just need to delete the API key over here. There we go. Boom. So that's how to use MCPS inside Quinn as well. How to code with it, how to use API, how to use image generation, video generation, etc. Now what we can do is we can compare these models side by side. Right? So if we go in, we've already compared the image breakdown, but let's compare for example like the outputs. So we'll compare DeepSseek versus chat GPT versus Quen. If we go inside the AR profit boardroom, we can just grab a example prompt. Let's go to the classroom here. Take the SOPs and then we can also compare this versus Gemini 2. 5 Pro. So I'm going to say, can you create a simple 3D car simulator with 3GS in a single HTML? We'll plug that into Quen. remove the image generation bit. Plug that in. Allow thinking. We'll just go with like medium on the token limit. And then we'll do the same inside chat GPT using 03. And then finally, we'll go over to AI Studio and we'll select 2. 5 Pro preview in the top right here. So this coding out right here it does honestly if you compare the interface the UI is so similar right like it feels like deep sea almost chat GP03 thinking about life here and that's creating the HTML and then AI studio is working its magic as well using Gemini 2. 5 Pro to be fair when you compare all the models Quen is by far the slowest because it's using that thinking mode but it's all about the output, my friends. So, let's click on preview here. See if we can use it. And that has created a bug. Let's see. We got AI Studio here as well. Boom. Chat GPT03. That's insane. Look at that. That is super impressive to be fair. Damn. Wow. 03 killing it. Let's have a look at Gemini 2. 5 Pro. That's doing its magic right here. We'll Can we run this on preview? probably can if we use advanced, but I'll just use livewe for now. You got to say that look I'm just in admiration of how good that looks. Obviously, you want to change the car a little bit, but for a first time coding, that's not bad at all. So, let's plug this in right here. This is more like what we're looking for, right? So, actually got a car, not just a big red block. I would say Gemini 2. 5 Pro definitely looks better, right? You can see the speed of that. It's like a PlayStation 2 game. And that was in one prompt, like one shot. Now, let's compare that to the output from McQueen. So, you can actually run this in artifacts. If you click the artifacts, but the preview fails. So, just something to bear in mind there. But what we can do is I can copy the HTML and just see if this actually works. Bear in mind it was the slowest. So, it used a lot of reasoning power. And then we'll take that. Delete that. Plug that in. And you're getting nothing. Okay, great. Let's see. Has he given us that's not working. So if you're comparing this side by side, like so far 03 pretty good. 5 Pro made the best output, which you can see right here. And then Quen is struggling the most, but we'll see when we get back in a sec. It's loading the preview. Is it going to work? Total fail. All right. So, from what I can see, so far, Gemini is still winning. It's still beating the outputs. If we compare, if we actually run this on live, we've actually seems to work, but it doesn't look good. Let's test that out again. Yeah, that's not looking good in my opinion. If we run the tests, it's like right, Gemini 2. 5 Pro smashed it. Chat GPT03 came in second. I really like the UI and everything of this. It's super nice. I just can't seem to I can control the car as well, but the car is just a red block, which doesn't make much sense, right? So, we can drive around and that sort of thing. And then if we have a look at Quen's output, it's not looking that great, is it? It's not looking amazing. What I can say is make BY, more beautiful, improve the car design, improve the UI overall. We'll see if we get better outputs from that. And also for me, like the preview is not working right, we can add Oh, we can add this as well. So, let's put this inside the prompt here. We'll refresh the Quen preview. Now, I'm just going to plug that in. And then it also says add this before the body. Okay. But it's not looking great, is it? Let's be honest. All right. We'll give it one more shot. Quinn says, "This highlights that even state-of-the-art AI models have different performance characteristics depending on the specific challenge. " Yeah, I would agree with that. Like for example, when it comes to UI, Claude is the best. coding stuff in one shot, it seems like Gemini 2. 5 Pro is usually the best. And then Quen, it might be like a state-of-the-art model, but it is struggling on this gaming challenge. Whereas, if you needed to create a video in one click, Quen will probably it's going to outperform all the models cuz most of them don't do it. So, let's take the HTML file here. Again, the preview is just not working. So, just be careful of that. It seems like I can't run 3JS or something like that. Now we've got the output, but it's still weird, isn't it? Look at that. What even is that? It's actually hurting my eyes. Let's refresh that. So, this is this has been corrected using the most powerful model from Gwen. And it's not great at all, right? Super average. Not even average. I would say that's worse. So again, like Gemini 2. 5 Pro, it's working a lot better. If we ran that same prompt through Claude, I can almost guarantee that it would get it right first time around. Let's just check that out. So go inside Claude here. Run exactly the same prompt. It's a lot faster to respond as well, Claude. Claude is going on a rampage right now with the coding. Look at that. It's just blasting it out. What we'll do in the meantime is just test this out for something else. So what we got here? This is some interesting prompts we got. Create a personal portfolio website. Make a small Gimoku game. All right. So, we'll try something else now. Say, create a beautifully designed snow day calculator. SEO optimized for the keyword of actual content on the page. Blah blah. We'll test that one out. We'll do that inside Gemini as well. I'm going to use Gemini Advance for this just because then we can actually preview it. By the way, actually, to be fair, Gemini has a video model as well, which is AO2. You can select it down there. Maybe we'll come on to that a second. I'm going to enable canvas here. I actually really like the canvas mode inside Gemini. Normally, it looks super nice. We'll see how that performs. And then we'll do the same inside chat GPT as well. Allow the canvas and see how they perform. All right, Claude has struggled. It's going on an absolute rampage. It's used up its own context. You see, it's reached the maximum length for a message. It's just going on and on like a friend that never stops talking. All right, Claude has not actually done so well there. Gemini 2. 5 Pro has performed the best, followed by Chat GPT. Claude has come out last. At least Quen gave us something even though it's not very good. This is the output from Quen. Feels like an old chat GPT style method. Let's just check is this working. So, I'm going put 1989. Then we'll put 20. Calculate snow day. Does the calculator work? No. Okay. This calculator is not working. Let's see what Gemini did. That's what I'm talking about. Gemini just it doesn't let you down. Look how good that is. Like the UI, the design, even the emojis in the title. It's just a beautiful touch. And that's using 2. 5 Pro Experimental. And then let's see if the calculator actually works. Boom. Yeah, it does. So, I mean, there's a huge difference still in Quen versus Gemini. Like, when you compare the two models for sure, like there no hype. Like, Gemini 2. 5 Pro is just destroying everyone right now. Let's check chat GPT. We'll preview this bad boy. Allow. Allow. Okay. It's not bad. It's not It's definitely better than Quen on Chat GBT, but it's not as good as Gemini, right? So, Gemini is just absolutely destroying everyone. Also, you can host on a subdomain if you want to share the canvas. So, overall, I'm still saying that Gemini 2. 5 Pro is absolutely smashing it. Chat

cheapy, I wouldn't switch from 03 to Quen anytime soon. And then Quen, it can do the job if you want something cheap and cheerful and it's great for that free API. But it's from what I've seen nowhere near at the same level as Gemini. And this is using the most powerful dense model which is Quen 32B. So actually when you compare them side by side bear in mind this is the most powerful model that I'm using Quen 32B. When you compare them side by side like there's not even a comparison here. Like the UI, the design, the functionality of the tool, the quality of the code, etc. It's nowhere near on the same level as Gemini 2. 5 Pro sadly, but I hope one day it does because it's a great model. It's a great idea. So, thanks so much for watching. What I'm actually going to do is I'm going to turn the whole of this video into an SAP. If you want all my best tips, templates, workflows, etc., you can get that inside the AI profit boardroom. We update it every single day, including if you like MCP stuff and oneclick agents, we've got a full SAP on this. Plus, we've got so many tools in here. Whatever you want to do with AI, you will learn how to do it inside the AI profit boardroom. Also, what you can do here is just type in anything that you want and you'll find it, right? So, for example, if you're like, okay, I need some help on building a website. Type in website and then you can see our whole collection on how we build all this stuff, right? And also, we do weekly Q& As inside here. So, you can just jump on the calls if you ever get stuck, if you want to learn stuff, you can inside the weekly Q& A calls here. There's actually one tomorrow if you join today. And additionally inside the community here, you can ask any questions you have and then you can learn from our members, right? So you can post in the community. You can see for example all these questions, all these community posts that we've got right here. Tons of great stuff inside there. So if you don't make the Q& A, you can just watch them back. That's just a tip of the iceberg, the tippy tip. All right, because we've also got the SAPs, the crash course, the Q& As's, the NA10 templates, recommended watches, ASO automation, AI agents, just like a redonkulous amount of content right there that can help you. So, make sure that you check that out. And also, if you want to get a free one-to-one SEO strategy session that shows you how to grow your business, get more free leads, traffic, and sales with SEO, feel free to get that inside the AI profit boardroom. All right. And essentially what this is going to do is give you a customtailored game plan to grow your business with SEO based on what's working for us and our happy clients. So if you want to learn all the stuff that's working right now based on we've got over a 100 testimonials now from happy clients then feel free to just get in a free strategy call. You can ask any questions. We can look at your website and we are happy to help you. All right. Thanks so much for watching. Appreciate it and I will see you on the next one. Butter.