From Nowhere to #1? How Qwen 2.5 Just Beat Top AI Models

Foreign. We are back with another AI model, Open source, free and better than Deep seat, better than ChatGPT, better than Clara 3.5. Here is everything you need to know about Alibaba's latest model, Kwan 2.5, including all of the hot takes, what this really means for the AI industry and how you can start to use it today. And is it really comparable to O1? I'll give you an exact prompt and I will show you the side by side comparison to see if it is worth using for your work. And did I mention it is free. All of that and more on this episode of Marketing against the Green. I'm Kieran, your co host. Let's get into today's show. Okay, we have another open source model from China, this time Alibaba, and it's called Quan. So we are just trying to digest all of the news from Deep seq. I just did a video breaking down Deep Seek. What is it? How we're going to how you use it, how it's comparable to the ChatGPT OpenAI 01 models, including their reasoning models. Right. It was a really huge deal because this model came out of nowhere. It cost a fraction of the price of OpenAI models. It's actually free, it's open source, so we can see how they built it. And it was a big moment within the kind of AI space. Right. All of the AI chip providers, their stock got crushed. Nvidia and all these other chip providers, their stock got crushed. Why? Because Deep Seek said, hey, we trained on way fewer chips than all of these American models and we've given you just the equivalent kind of power. And so everyone's trying to, everyone's been trying to like wrapped their heads around Deep sea, this deep seq R1 model, open source from China and that was about a week ago and that has been the talk of the Internet. Their mobile app was number one in App Store, which is actually pretty incredible when you think about it. It really went viral. And just as we're trying to get our heads around what this means for the AI space, we have a new model from Alibaba. We have Quin 2.5 and guess what? It's it is equivalent, if not better than Deepseek, than ChatGPT's own models, then Claude 3.5 Sonnet and it's another free open source model. You can go to the website today and sign up and to start using it. In this video I want to give you some of the chat that's been happening online around this model's capabilities, some examples of it in action. And I want to test it with one of the kind of advanced prompts that I use to basically do my work against Zero One Pro, which is ChatGPT's $200 a month model. So you can see it actually in action and see a real comparable. But let's kind of jump into some of the news here. Okay, here's one of the kind of breakdowns here. It's from Ruben Hasad and he has some really great takes on some really great updates on AI. He is the founder of EasyGen I. And so he talks a lot about here, kind of the breakdown of Quinn 2.5 from Alibaba. You know, it can code, it can write text, it can search the web. The big thing is it's multimodal, right? It can generate images, it can even generate videos. And so we're going to get into a little bit of that. Some of the things maybe just to call out in terms of what he talks about. And I won't go into the technicalities as much as I'll just go into the, the actual use cases and some of the interesting things you can do with it. It actually is able to handle a lot of different documentation and so that seems kind of a small thing, but it's actually a pretty big thing. Some of the other models I'm looking at Claude specifically, but even OpenAI's O1 reasoning model cannot handle a whole range of different documentations and this one claims to be able to handle and parse many more different types of documentations. It can actually handle has large context windows, can handle large amount of video, it's open source and so we're going to get through a little bit of that. But like basically you can go and actually very similar to Deep Sea go and basically look to see how this is built and people can kind of quickly build on top of it. It talks about how it has been trained. All of that is contained within the open source papers. But this is what I wanted to talk about a little bit. So Quinn 2.5 max is the red bar here and these are the different benchmarks that models are graded against. And you can see deep seat V3 is the blue and then you have llama, you have GPT4, you have Cloud 3.5. So you can see in some cases it's performing on par, if not even better on some of these benchmarks. And if you look at Kwan and Deep Seq together, they're pretty equivalent to these paid models from these U.S. companies. And I think that is starting to blow people's brains, which is while these models are our free models, we can start to use them. And it puts a lot of pressure on the other AI companies who are currently charging for equivalent access to equivalent models. And I want to get through, I want to go through what that really means at the end of this video. Thing I've learned about AI is that it works best when you treat it like a collaborator, not just a tool. It's there to help you get moving. It sparks ideas and it makes starting just about anything 100 times easier. Now, to help you tap into the potential of AI, we've worked with our friends at Mindstream, an amazing AI newsletter. We've teamed up with them to release 1000 expert level AI prompts for everything marketing you might ever imagine. When you need some fresh ideas, when you need to adjust your strategies, content creation, inspiration. These prompts are designed to make starting and staying creative a whole lot easier. Get it right now at the link in the description below. Now back to today's show. Okay, I wanted to go into one more kind of thread which I thought was a pretty great breakdown when I was going through this early my morning today. And so this is obviously you have to start with a hot take people. If you're going to do content you need to have that hot take. I do think that you the kind of hot take ification of everything is a bad thing. I would love to be able to get back to just some normality and content but you have to do that today. But actually the content here is really good so talks about the fact that the Q we're going to get into actually run across and look at these different models because there's a bunch of different models you can pick when you go and sign up for an account of in Kuon. Probably pronouncing this pretty wrong I would say, but everyone knows who really watches this podcast. It's pronunciation is like my worst skillset and so it's a train on the massive amount of data. It uses special techniques to learn better and performs as well as top. And so again in a future episode I might get into some of the specifics and how they train. This model they do have. It is open source so I do think you can go read about this today. I'm coming at you live after looking at this this morning. And so some of this will have to be follow up episodes because I haven't dug into the papers yet. But it does beat V3. We looked at some of these benchmarks in the arena Hard Live benchmark and GP QA diamond benchmark. So we kind of looked at those benchmarks just previously and you could see again that it performed on par with other models and in some cases was better. In Deepseek the models are available in Huggins face and Model Scope. Again they are kind of open source. And so one of the things I said when I did the video on Deep Seq is I think this is going to increase the pace of evolution in terms of custom models that are available for people because now you actually have these great foundational models you can just quickly build on and craft for your audience. And I do think that companies are going to start to take a bunch of these open source models and actually add propriety data priority ways to train them for their audience and have like custom models. I think that's really important because it does make me believe and I have always believed that we are going to see companies use a whole mix of different models including custom models and a lot of open source models to do their work. And that's going to put pressure on some of the conglomerates because you may not have just like one single model like an OpenAI because that doesn't is not the best model for all of the different tasks that you have. And so I do think what we're seeing is this evolution on how models will likely be used within companies. Talks a little bit here about the key AI test. So there's like a good len view here that we could probably quickly get into because it does, it's a, it's a pretty good snippet on how the, how it outperforms these other models. Matt Alibaba's just dropped their latest version of Quinn Quinn 2.5. So this is Deep Seek followed by Gwen. This is now a much bigger issue. It's not just about China us. This is about open source versus closed source. What are they promising at Alibaba? Yeah, that's right. So I mean I guess this kind of place point I just, just said that you know that the mark the market is evolving at a really rapid pace and we've seen this with the Kuan model. I think right now it's too early pronouncing it right after all to say. But obviously the headlines from Alibaba this new model outperforms the Deep Seq model across a whole range of kind of industry standard benchmarks. There's a lot of gaming of these benchmarks so I think we'll have to wait perhaps a matter of hours or days Whilst engineers get their hands on this new Alibaba mod performs better than Deepsteak a bunch across a bunch of these benchmarks that we've said before. The interesting thing is he said something. Well, he said something very interesting there which is like some of this stuff is gamed and that is there are numerous takes. You can go read on that. But there is some validity to the fact that he that there is ways that AI companies can actually game if you pre train on the data that you're actually being tested on. Now I don't, I won't go into that because I don't know enough about to have a really hard point of view but I think it's going to have to play out over time how good these models really are and comparable to others. If you saw Sam Altman's quotes recently, he would tell you that OpenAI are still much further ahead in terms of their models and the models they're going to release. But he specifically said Sam in a tweet yesterday that they were going to rapidly release their models quicker than they had originally planned because of the pressure they're being put on there by these companies. And that comes back to what I said, which is we're going to. The pace of evolution in the sophistication of models is going to just go faster. It's going to go faster and I think we're going to see the large companies come out with models much faster than they had planned and we're going to see companies start to fine tune and customize models much more than we had thought. Go through some of the use cases that I'll go through but like much, much of them are equivalent. If you've used 01 Pro or even Claude 3.5 Sonnet it's going to feel pretty much in terms of the things that it's really good at. And again that it's a free model, it can create like some really great images. I want to get into that because it's Dolly competitor is really actually good and surprisingly good. But it also has a Sora competitor because it has video as well. Let's get into the actual images. I will just stop here because it does, I have noticed as it can scan and synthesize information better than most AI models like it is able to. That's one of the things that I'm going to show you. One of the use cases it is able to really have like clear and concise takeaways from large amounts of data. From the small amount that I have managed to Use it this, this morning. Here is one of the videos she generated. I don't know how people are generating videos because I don't have access to videos yet. So I don't know if it's because of my European status, but I can't get video. I can get images. But here's a nice example of the video. It's not bad. Like a guy traded Nvidia stock on his computer. You know, the really annoying thing is I was going to buy Alibaba stock and then they came out with this and there's no point. So that sucks. All right, so here's, here's one of the videos here. I think it's like pretty good, right? Like pretty equivalent to Sora. I actually think it's. Now I will say when you do the nature, like the realistic Sora, when you ask like these video models to do some sort of nature or anything like that, they usually look really good. Where they go really bad is where you have different objects having to interact with each other. And there's some well known limitations of AI video models of why that happens. That again I won't go into because I'm not the most technical person to describe why that happens. But I think that's a pretty good example. There's other examples. I will show you one more on the video. One more. So it's. This would be a quantify plus. Oh, I've got the wrong model chosen. That's what it is. Let's go back over and I'll generate a little model, a little video. So it's really slow but free. Free. A detailed and hyper realistic scene from a classic King Kong movie. But a massive kitten as Kong terrorizing downtown. So like here we are. Pretty kind of cool, right? Like it's pretty good quality. I don't know. This one here is pretty good quality as well. So I has like some fun AI tools, some fun imagery image tools as well. The other thing I just wanted to go before we get into the model itself is if you're looking to like what can this model do? They release these cookbooks. These cookbooks are kind of cool. So they release these kind of cookbooks, right? People just talk about the performance but often pay not enough attention to how the model really works. I think this is really great that they're releasing this because I actually want to go through these. Which is what I was saying is like I will probably do a follow up episode with some of the core technicalities because I have not managed to go through these cookbooks yet. But to help you understand the model, they've made these kind of cookbooks for you. The cookbook includes notebooks for multiple usages. So it actually gives you ideas of how to use the models, including computer use. So they do have the ability to do something similar to like a Claude or OpenAI's operator, where they can actually build something like an agent to use your computer. They have a mobile agent which we can get into. They have video and they have like this really we talked a little bit that in the, the video understanding the ability to actually understand long form video. So you can ingest a bunch of video and it can understand it and do kind of clear, concise takeaways or you can interact with that video in interesting ways. So the cookbooks are basically these things, right? So like they have different cookbooks and suggestions on how to actually use the model and how the model worked. And so the computer use agent, I wanted to quickly show you this one because computer Use Agent has been the talk of the town because it's really the automation of work. And OpenAI's launch probably got the most traction, which is OpenAI's operator, which is an agent that can do things for you, can operate, can complete tasks when in a browser. And so if you see here another example of the strength of these open source models and how much pressure they put on these kind of large AI companies that are trying to charge for a bunch of these features. Basically Quan 2.5 VL lets you create a $200 a month OpenAI operator like Agent for computer and mobile phones. So two, but two of the cookbooks are the computer use, which is the ability to do things for you and the mobile agent that that comes with this model, which means you can use it on mobile as well. It's 100% open source. And you can see this person here has built a pretty cool agent to be able to go do things on booking.com go book a flight. So very similar to OpenAI's operator. But for OpenAI's operator you have to spend $200 a month for this here. You can easily use it for free. And I think that is huge, right? Deep Seek have something very similar. People have been building agents in a similar way for Deep Seq. But Deep Seek obviously has a ton of privacy concerns about it. I have not read into the similar, I've not read into like what the privacy is for Alibaba or where they're sending your data. But again you can start to use these open source models to replicate a lot of the things that other companies are charging for. All right, lastly, let's just quickly show the model in action. And so let's go over. All right, so let me show you basically what I was doing here. All right, so this is a PDF. Basically this is a simple use case which is it's a growth dashboard for a SaaS brand. That's PLG. And what it's going to do is help me figure out insights and then put together a plan of how I can continue to grow this business. It's going to isolate the key metrics, look for bottlenecks, and then provide a strategy to actually solve those problems and grow those metrics. So I use this, I've been using this because I'm building a little course that I'm doing a free webinar tomorrow to kind of show the capabilities here. And so I've been using this with O1 Pro, which is the 200amonth from ChatGPT. And so let's go up so we can show the equivalence. And here you see, I can have the dashboard and I basically, this is a prompt. It looks nicer in my Google Doc. Obviously it's unformatted here, but there's a large prompt. And the reason I say a reason and prompt is because the difference between a reason and a non reason and prompt is the non reason prompt is the traditional kind of prompt where you say you're an analyst. Here's the step by step thing you need to do. This is exactly the stuff that I want for a reason and model. You don't do that because reason has the step by step logic built into the model itself. Really what you want to do is provide it with the problem you're trying to solve and the output you kind of want, but let it try to figure out how it's going to do that for you. And so this is a reason and prompt to figure out. This is much more tailored for reason and engines. And what it does here is it goes through and looks for the metrics that I should focus on if I want to continue to grow this business at 100% year on year. So I want to double ARR. And ChatGPT01Pro does a very good job here where it will go through, evaluate all of the metrics and tell me like what are the most important trends that I should be aware of? And then it actually tells me the metrics that I should prioritize to actually hit that number. So I don't want to work on all the metrics. I want to work on the most important so that's the first part of this, the use case here. The second is to build a forecast model. The third is to actually build a strategy to hit that forecast model. But it's a good equivalence because I do the same thing in quan. I kind of give it the growth dashboard and then I ask it the same sort of thing and it does a really good job. Again, this is a free model so it kind of breaks down the kind of key growth levers. It's kind of very similar. It talks first of all just about the importance of these metrics. So it evaluates acquisition priority analysis, talks about activation and onboarding priority analysis, talks about monetization and expansion revenue priority analysis, retention and churn very similar viral so it talks about each of the ones and you know it's kind of interesting. It's actually got the exact same core metrics as ChatGPT to prioritize has it. Let's just see this is kind of interesting. Yeah conversion rate. Yes. Churn yes. RP user retention rate, referral rate and activation rate. So it actually deviated for these two has referral rate but doesn't have user retention rate. So it actually the chain this is the one that shows activation versus retention which is interesting reasoning impact. Let's see. Ensure more user experience core value quickly boosting accelerates the path from sign up to paying customers enhancing both activation and conversion metrics. So it does a pretty good job. Same here again this is the first don't expect it to have the exact strategy because I have not started to prompt that yet. It does a pretty good job of like isolating the core metrics and then I would ask it to build a forecast model of what these need to be for me to continue to grow 100% so I found that the kind of prompts that I've done this morning, the kind of reasonable prompts where I've asked it worked here with strategic problem is somewhat equivalent to ChatGPT Pro01 Again this is like ad hoc some prompts not all the things I'm sure there are cases where they deviate a lot but for me it's like pretty similar and that's not even that was just a capable of complex that's just their plus model. They also have the MAX model which is their most powerful so maybe I'll rerun that where I'm showing you other I could rerun it in max. Can you rerun this analysis and we'll do a new chat warning. Leaving this chat will interrupt the current generation, it cannot be resumed. Oh, interesting. You can't do multiple chats at once. Oh, geez. It's gonna start to do some pretty dope things in it. Okay. This is. The more I have not used, the more powerful. I've not used all of the models because again, I got access to this this morning, which is cool. Me being in Europe, I can actually feel like I have access to things like everyone else. And so I think it's gonna pretty like they could, it could be the reason. And the, the really strong models you really have to have like, like they have to be quite complex to the point where I would think of it as when do I need to use a complex reason? The model usually when there's like five or more steps that model has to run through to complete it to complete the objective or goal that you have. And so it could be that the one that I'm given at this is going to look very similar to the Q 2.5 plus because it's really doing analysis and then pulling out core metrics. I think if I had asked it to build the strategy and then create a two page memo as the output, it would have to run through multiple different steps and it would probably delay the. The output from plus and Max would look a lot different because this is a more powerful model. The more powerful model, the more strategic and more steps it has to run through, the better it's going to be than the model that is like less complex. Right. And so they're probably. The output will deviate a lot. Whereas I expect this might actually look kind of similar to the model I had just used. Let's just see if it's got the same thing. Yeah, it's like very, very similar. Right, okay. It's giving me strategic. But anyway, I will say that these models, O1 Pro or maybe Quan because I have to get again, are pretty amazing to do. Like first of this stuff. If you have like large amounts of unstructured data, you want to find the core levers and bottlenecks in there, you want to build a forecast model, then you want to give you strategic insights into what are the things you should do to continue to solve those core bottlenecks. As a first pass, you still have to do work after that. It's pretty incredible. So the other thing I wanted to look at is an image of a dog having beer after a hard day's work. I mean, cold beer, half day work. I don't drink, but I'm sure I used to cold beer after a Hard day works pretty sweet. Let's just see how good the images. There you go. I'm, you know, it's. The image too is pretty cool. From everything I've seen online, it's pretty cool. I don't know why I don't have access to the video. I'm actually really interested in going to do that. But I've access to upload all the documents, everything else, else. So this is a quick introduction to quantum 2.5 A. And so Lee, I'll leave with this. What has happened over the past week? Well, I think what's happened over the past week is the money has moved from potentially AI models and hardware. Although I would put question mark against hardware, which is Nvidia and these companies that supply the chips, I think their stock is going to go back to where it was. But it's moved to maybe, but it's moved to the wrapper layer. And originally when the first wave of AI, everyone I think also including me, kind of said these rapper companies are not going to be big companies because they're just templatizing use cases for an audience and just kind of wrapping those templates around these models. And these models eventually will just be able to do all that stuff for you. So they'll just kill the wrapper companies. And I think that's probably true of the very first wave of rapper companies that they just were not complex enough, sophisticated enough, and I think we believe be included that all of the money would probably be in with these conglomerate models because the models could do all of these use cases for people. What we've seen over the past week is that is not true. It is the model part that is being commoditized. We have two incredible models that are now free. Obviously they have privacy concerns. I have not looked Deepseek. You would want to run locally on your machine. So it's not sending anything back to the core servers. I have not looked into Alibaba, but you can even get Deep Sync through perplexity to start using the IT today. So there are privacy and privacy and security reasons. You might not be able to use these models within your kind of enterprise use cases, but they are an example that the model layer is being commoditized and these free models are putting huge pressure on conglomerates. The conglomerates will have to release models quicker for cheaper and probably provide them for free. Like OpenAI, I think has already realized because they're giving away a bunch of the O1 stuff for free. That means the problem to solve is just getting people to use AI and have it be impactful for them. So it's really the use case and how you get it into people's daily habits and workflows. What I see in AI all the time is people think it's magic for the first week, the second week, oh, it's fine, I'll use it now and again. Third, fourth week, they don't use it anymore. That is the hardest thing to actually solve is integrated AI in a way where it becomes very useful for people and actually impactful for people. And user habits and workflow changes are the hardest things to actually change. Right. And so if you really want people using these models, you have to figure out how to integrate it into their work on a day to day basis. And I think what we're going to see is a huge increase in the app layer. The app layer, the apps getting built around AI. And that is where all the money is going to accrue. And that is going to lead to these app companies, small app companies, becoming huge businesses. I talked to a founder recently multiple times. One of the fastest growing companies in Europe. Tiny team, growing at eye water and pace. Right. And this is the new kind of startup we're going to see. They're, they're in the app layer, they're doing something really smart in the app layer and making AI easy for people to actually consume and use. And they're growing at record pace like one of the fastest growing European companies of all time. This is where the money is going. This is what the core takeaway is over from, from the last week. Play around with Kwan. I hope this was useful to get started. We'll be back with some better use cases when we get to dig into it a little bit more. But I hope this was a good jumping off point. This data is wrong every freaking time. Have you heard of HubSpot? HubSpot is a CRM platform where everything is fully integrated. Whoa. I can see the client's whole history. Calls, support, tickets, emails. And here's a task from three days ago. I totally missed HubSpot grow better.