DEEPSEEK V4 + OPENCODE + SUPERPOWERS IS ABSOLUTELY INSANE

Hello you beautiful people and welcome to this video. Now I was not expecting to make this video today because I went to bed and GBT 5.5 was the fresh hot news. Now I wake up and I mean since then I mean look at this. All of these models have dropped since then. That is absurd by the way. But obviously the big news Deepseek V4 Flash and Deepseek V4 Pro have dropped. Now I think I will do a test on the flash model today. The reason being is I prefer personally cheap models. This is a huge step forwards. Now a lot of people might say what are you talking about Hish? But if you noticed last time this happened that is when all of the really really good Chinese models started to pop off. I have a theory. I I don't know if it's true or whatever but I think Miniax and all of those other models Kimmy K blah blah blah. I think they're all just deepseek models that have been trained in, you know, slightly different ways. Now, take that with a pinch of salt. I'm not an expert. You would expect me to be an expert by now, but that's just, you know, my that's the kind of thing that I've got like the impression that I've got is that all of these models, the other models are just deep seat models. So, what does that mean? It means probably going forwards the open source models, Kim K, you know, all of that good stuff will be 1 million context window. So, that's why this is such an important piece of news. I've done some research into this model already, and I've heard that it's not incredible, but I don't think that's the point of this. The point of this is that from now on, the new models going forwards will be absolutely amazing. Now, I'm going to do something that I don't think I've ever done on this channel. I'm going to give you guys a music suggestion. All right, I'm absolutely obsessed with these guys. I'll leave a link in the description of this video. They're called New Jane, which I think means like new generation. They're a Neapolitan funk and soul house kind of vibes band and I'm telling you guys this will be the this album here will be the album for your summer if you go and listen to them. Couple of really really good songs guys. This one here Madaka um number four on this album by Mediterranean. Absolutely incredible. This other song here shala as well absolutely incredible guys go and listen to this stuff. You'll love it. I'll leave a link in the description of the video. You'll be blasting that out on your speakers all summer long and you'll be saying thank you to me after you've listened to it, guys. It's honestly absolutely banging. They're about to release a new album as well. Not sponsored, by the way. I just thought I'd mention that. I'm probably going to mention it a few more times because I'm absolutely obsessed with that band right now. And you know, why the hell not? Doesn't have to be always a Harbor sponsor. I can just talk about the music that I'm listening to as well. So, let's dive into things. First of all, I just need to work out how we're actually going to do this test. I'm going to guess it's going to be open code. Um, but I'll be back in two seconds. Okay, so I'm not going to really be looking at the the benchmarks and things. I'm not really too bothered about benchmarks, honestly. I've I've stopped looking at them. Like, I really just don't care about benchmarks. They after GPT5 claimed to be better than um Opus, oh, sorry, GPT 5.4 claimed to be better than Opus 4.7, I just knew that benchmarks were cooked. I just knew it. Okay, so we should be good to use open router here. If you'd rather use open uh deepsek direct API, no, that's fine. Okay, and then we'll do mk dear deep seek v4 cd deep seek. Can't even write this morning. V4 then uh open code. Let's see if this instantly gives me no. Oh, there it is. Beautiful. It's already available. Okay, so we'll use flash here. I'm very, very curious to see the result. Let's just grab the prompt. By the way, guys, it's absolutely beautiful in Ireland today. It's 20°, which like if you know anything about Ireland, that is mad. That is mad for Ireland. Yeah. So, it's 18° right now. Let's just see what it is in Naples, where I used to live. Naples weather. It's going to be well hot, right? Oh, it's not even that hot. What? Scammed, bro. I thought it'd be like 35 degrees today. I guess it's still April. Okay, so we have to allow our data to be sent to China happily. I really don't care. Like honestly, people are always like, I don't want to give my data to China. I don't want to give my data to anyone, let alone like it doesn't matter to me that it's China. I don't want it to go to chat GP. I don't want it to go to anyone. But like we're we're too deep now, guys. We're just too deep. So like there we go. It works. Beautiful. Wow, this is super fast. First of all, uh don't don't use skills. Don't use skills. Don't use skills. I do like the fact that it um it tries to use skills. That is a good sign. Okay, so here we go. It's starting to read. Directory is empty. Empty. Beautiful. Let me plan this out first of all. Let's check if no JS is available. Beautiful. So, look at that. This is actually It's not better than but it does like I I don't know what it is with GPT, but it always just tries to um go into a different directory, right? Which is clearly not what I want at all. So, let's see. The instructor is waiting. Beautiful. Let me try with a yes. No. Okay. A couple of things to note about this guys is this um is absolutely not possible to run on like a local machine, right? Unless you have an absolute bloody beast. Okay. So, let's see if this can actually do anything here. Let's see if there's any information. Deep Seek V4. I I don't want to look at benchmarks. I really cannot be asked for benchmarks. Um but I do want to see. Okay. So I always get sucked into benchmarks anyway after saying that. So Opus 4.6 is the this one I think. Yeah. Okay. So like I said guys, I am a pretty big fan of flash models just generally speaking. I've always been a fan of flash models. They're cheaper, they're faster, and I just like them to be honest with you. One of the problems is this is what like some crazy big model. Um I don't actually know how big they are. So, 284 billion parameters. So, yeah, you're not going to get this running on a local machine or anything like that. You would need an absolute beast. You guys might say, why are you not using the Pro model? I just I really really like um flash models, right? That's why I'm using the flash model. If the flash model can't even do anything, then the Pro model is probably going to be crap as well. This is nice. 1 mil context is super nice. That is a huge leap forwards. If you look at all of the not Kimmy Oh, yeah. 2.6. If you look at all the Chinese models, for example, I think they're all uh 250, right? Or whatever. Let's see. Kimmy 2.6. What is Kimmy 2.6? Am I wrong? It's Kimmy a million. No, 256. So, if you actually look at Deep Seek, is it was it V3 the big release that they did? I guess this is also okay. 163,000. So I like I said I don't have proof of this but I've always thought that these models are just deepseek like trained in a different way right so this is a huge leap forwards for us because this now means that um all of the models going forwards will be a million. I don't know what's going on here. It just stops building. Yeah. So I'm not sure why this is just stopping every single time. I've never seen this before. I might have to hop on over to the pro model just because this is pretty much unusable. What this often is is a prompting gap. Change to this model. So what I mean by a prompting gap is like the model isn't uh sorry the prompts inside open code aren't fit to the model. Right? It could be that it's not good enough or it could just be a prompting gap. Right? So, before we completely sack off this uh this deep sea model, which by the way, I don't think this is going to be good. I I I'm like pretty confident this model is not going to be good. That's not the point of this model, though. The point of this model is to bring all the other models up to scratch, right? So, like all of the recent good Chinese models, the whole point of Deep Seek is to give them a better base. Okay, I'm getting rate limited on V4 Pro. Beautiful. So, not the best of starts for me. Um, can't use the flash one cuz it's not good enough. And Deep Seek V4 Pro is already getting rate limited. So, RIP. Looks like I woke up too late today. Jesus, it's 2:00. What the hell is wrong with me? Okay, let's see if we can do this with the official API. RIP just leaked my API key. It doesn't really matter. I've only got like 30 cents on this thing anyway. Okay, let's see if this one can run this. I might have to top up. I don't really care. I'll just top up in the end. It does already have a project to work on, so that's fine. That was just from the previous Flash thing. By the way, guys, I'm not completely on Flash just yet. Although, it does seem like that was not a very good run, I have to say, which is a little bit disappointing. But it might just be that the open code prompt does not fit with Deepseek Flash. So, I want to test a few more things. So, I'm just going to do a couple of things like create me an SVG of a birthday icon. Let's just see what it does here. And then we're going to let this one run. I'll top up if I have to. I'll give all my data to China. I don't really care, honestly. I know I'm going to get a flame for that. And someone will probably call me a what is it? Like a what's the word? Someone will call me like an agent for sure. But like I know that there's someone watching this right now thinking, "Yeah, that's exactly what an agent would say." I'm not paid by anyone, guys. Someone said yesterday that I was paid by Anthropic. I'm not paid by Anthropic just for saying that they're better than Chat GBT. Like, it's embarrassing that people actually think that. I wish I was paid by Anthropic. Like, I would I would accept their money 100%. I would accept their money because they're, you know, I've talked about them since day one pretty much since they were since they came onto the market pretty much. told people that Claude is good. Um, and then, you know, since Claude Code was released, I've always talked about how good Claude Code is in my opinion. But yeah, I mean, I get why people think that I'm being paid by AI companies. I wish I was, but I'm absolutely not. This is a pretty good result. It's not actually bad. Okay, so it actually asked me a question, which is rare. Um, normally these kind of models don't. It's still asking me questions. So, generate SVGs. Beautiful. That's a really good sign, I have to say. That's a really, really good sign. I don't know if this is open code changing whatever but use placeholders. This is not something that is very commonly done. You can see down here we have the cost so far. Uh I really don't have a lot of money on my DeepSeek API key. I will top up if I need to. But I'm just going to continue to do a few tests that maybe I haven't really done before uh on the channel just because I want to know if this model is half decent. Now like I keep saying in my opinion it doesn't actually matter how good this model is. it matters how good the models that are built from this model are right that's my opinion again I have no proof that Kimmy 2.6 six for example, it's just deepseek iterated upon a million times. But I do have that feeling and obviously China is trying to catch up in the AI race, right? Just makes sense. That's just the way things are that you nobody in the world wants America just dominating everything except Americans, right? I don't know what it is with Americans, by the way, but they seem to think that everyone seems to just prefer them to the rest of the world. And that it's just not true at all. Like I don't know what it is. Uh, it's really, really funny. Honestly, I'm gonna get flamed for saying that for sure. But luckily, not that many people watch my videos anymore, so I can kind of say say whatever the hell I want and not get too flamed for it. But yeah, I just find that interesting that like you talk to Americans and they just assume that you would rather they you know, you you would rather send your data to the US than another country. I mean, I wouldn't I don't really want to send my data to anyone, right? But at the end of the day, I've already sent all my data to everyone already. So in before Hamish is a Chinese agent comments in the comments. Go and comment that now guys. Okay. So this is just a simple uh test of what we're actually doing with Deep Seek. This is just a um like a one pager. Oh, that's not bad. Honestly, that is not bad. Damn. Yo, if it can do something like this for the actual result here, that is 100% not bad. I don't know why W3 schools had to add all these ads. I mean, obviously I know why because money is money is number one in life, but this is pretty good, honestly. I'm going to try a few more prompts here. So, um, just go here now. Create me a HTML, CSS, JavaScript, 3JS, uh, FPS shoot game. I'm just going to give it that as the prompt. No more info. Okay, this is good. Yes. Okay, so I mean these I'm just going to say it. This seems like a huge step forward for Deep Seek, which means the models that come after this are going to be really really good, right? Play fair display the classic. That is literally the font that is used by AI for the last two years. I I don't know. It's really really interesting that they always choose that always choose that font. It's always player. Why is it getting me to confirm everything as well? I'm really not used to this from like a Chinese model. Normally they just say okay and then just build some crap, but like this is really really really going into a lot of detail I have to say. And it's extremely cheap. We've only used 3%. Is this a competitor to Gemini 3 flash? Like actually that's very interesting to me. At this point I'm thinking about buying like a mega laptop. The problem is I want a Mac and it's like eight grand for the Mac I want. But I mean I would love to get like a Mac that would be capable of running this kind of model. Imagine you're on a flight and you can sit there coding with your own uh what is it like CPU or GPU or whatever the hell. That is so interesting to me honestly. Okay, let's see how this did. This is going to be very very interesting I would say. Where's the run? There it is. Okay. Why was that in Oh, right. Yeah, it was in Italian because of um I was writing in Italian before. Uh yeah. No, kind of. Why is the WSD not moving me but moving? That's so weird, right? No. Okay, this is terrible. Absolutely terrible. I can't even leave the game. Um RI. Hello. Okay. Uh no, this is not very good. But I mean, the prompt was pretty terrible. Just before we continue, guys, huge shout out from our sponsor, me. This is harbor seo.ai and it is my content generator that I've been working on for two years that is more than just a content generator now. It is now a full-blown SEO tool and my god the things that are coming to this tool are really really exciting. I have some insane plans for this and you really really want to go in and lock in the pricing now because the pricing will be going up soon. Whoever unpublished one of their pages, by the way, please go and publish it again because it was getting impressions. Now, this pricing is probably going to have to go up when we release harbor build. So, if you don't want to get caught by the pricing going up, go and lock in the pricing now. You can get a free trial for 3 days. As long as you are a paid member, I will never increase the pricing on you, right? I just won't. Whereas, when we do release Harvard Build and many other things, then the pricing will likely have to go up. This is not just a content generator anymore. It's a full-blown SEO tool, and it's helping people get real impressions and real clicks for their businesses. You know, almost 300 clicks a month for brand new content on people's websites. Right now, you can go and get 25 articles a month for €29 a month, which is absolutely absurd compared to all other content writers on the market, which are all way too expensive. I am going down the different route. I want to give people the best possible SEO, AEO, GEO tool at the best possible price, and that is my commitment to you guys. So, go and lock that pricing in, guys. Go and support the channel. There's a link in the description and in the pin comment. Let's jump back back into the Deep Seek content. Okay. Yeah. So, 3JS games is definitely not its uh strong point. That's okay, though. You know, these models don't have to be good at everything. I think the design for the homepage is pretty damn good, though. I have to say all Oh, this is Flash. Oh, damn. Wait, let me change to Pro up here. Okay, so let's just try the same prompt that we did before with um Deep Seek Flash or with Deep Seek Pro. What the hell is getting new chat? There we go. Jesus. Okay, so let's just do this. My cat is just bullying the dogs as usual. Okay, so it looks like Deep Seek Pro is fairly unusable um on Open Router. So I definitely if you want to test it out, I definitely would recommend using the official Deep Seek API if you don't mind, you know, signing up to a Chinese API provider. Okay, so interestingly, this did seem to use the superpower um thing in the end. I I didn't I asked it not to do that, but I mean it it happens, right? We've only used 5% context, though, so that means it's pretty much done nothing so far. Okay, so we actually got a run from Deepseek V4 Pro inside Open Router, which I didn't think would be possible today to be honest with you. So, we'll see how this runs. See how this does. Um, honestly, I'm going to say it. This isn't bad. Like, a lot of people are talking about it. I think people's expectations were that this had to be better than Quen 27 Bill for example. But the point is now Quen can build off what these guys have done, right? And make it even better. But honestly, I don't know what they've got in the water over there at Quen, but Quen 27 Bill looks like an insanely interesting model, right? If you know anything about open source, guys, this model here, there is something about this model. I don't know what it is, but it's super intelligent for super super cheap and also super small model, right? This can run on most computers. If you have a computer like even from 15, maybe not 15, 10 years ago, 5 years ago, whatever. If you have a desktop PC, right, this likely will run on that. Whether it will be fast or not is another matter, but at the end of the day, you know, you can't have everything. So, if it's a bit slow, but it's free to run, then obviously free to run is better than paid to run. Okay, so let's see what it came out with. This is the pro. This is undistinguishable from Opus. Literally undistinguishable or indistinguishable, I think, is actually the word. That's crazy. I know that this is just HTML, CSS, and JavaScript, so it's slightly different, but still like the design here is as good as, you know, the the big boy himself, right? I hate to even mention him, but yeah, I mean, that is absolutely crazy. This is just the file from here, just uh so everyone knows. It's not the build that I'm doing. That will be done, you know, pretty soon. It's taking quite a long time, but uh yeah, that's using the Superpowers plugin. So it'll be very very interesting to see the end result here guys. I don't know why people are so critical of models these days. Like this is a huge leap forward for DeepSeek, right? Which in turn is most likely going to be a huge leap forward for the open- source movement in general. Which by the way I do have to say the open source movement saying it like that is a little bit I don't know because it's basically just China trying to undermine the US by giving models for free to people. But still open source movement. You know, a lot of these models are probably originally based on deepseek, right? Again, I don't know that for a fact, but my instinct tells me that this is a huge result. Like this design here is really, really good. By the way, like just completely unrelated to what I'm doing right now in this video, but when the hell did they make Claude based? By the way, I haven't sworn to this guy in ages, and he just literally just me. That's so funny. Okay, so we have our first rumblings of completion here by DeepSeek. I have to say like open source is getting better and better every single test I do. This looks like is it Opus? No, it might not be Opus. This looks like some other model. I'm I'm not I'm I'm drawing a blank on which model. But very interesting stuff, guys. This is looking It hasn't finished, by the way, which is why it's not quite fully done in terms of the technical build. Um, I'm not going to like lie to people and say that this is the most amazing design I've ever seen. It's not bad though. It's literally not bad for such a It's not even cheap, but an open- source model, right? It's not even that cheap, to be honest with you, though. There are cheaper models, but honestly, fair play to Deep Seek. This is better than I was expecting. This is better than people are making out on YouTube. I don't know what it is with uh YouTube. It seems like there's a bias towards specifically towards Open AI. Like I'm just going to talk really frankly about this. I don't really see GPT 5.5 as being, you know, anywhere near 4.7 Opus levels, right? But anyone and everyone is just saying that it's like this incredible. I might be wrong. Maybe I'm wrong on that. Maybe I need to look into it some more. But um and then they just all over Deep Seek for some reason, saying it it's bad. This is from YouTube comments, by the way. I've been reading uh I've been watching videos on Deep Seek basically the entire morning. A lot of the YouTubers have saying that it's not great. A lot of comments are saying that it's not great. I'm not seeing it, guys. I honestly think this is probably the best open source model that we've ever seen um inside open code, right? Okay. Okay, so in insufficient balance, I'm not going to bother topping up. That only cost 0.46, though. It ran for I mean I I want to say a long long time. It did use the superpower skill, which by the way, you can just do superpower plug-in uh Claude code or whatever on um on Google and just find it, right? But yeah, I mean, I'm probably going to leave the video there, guys. But this is super impressive stuff. Um, anyone that tells you that this is pure trash probably has some kind of ulterior motive. Um, I don't at all. I'm not paid by any AI. Well, that's not really true actually. Z.AI paid me $3,000 for a video that I would have made anyway and I haven't made yet actually, but I need to make a video on their platform basically. Um, that is the only AI company that has ever sent me any money. Um, and yeah, I mean I I'll be transparent with people. I'm in talks with being sponsored. That's the YouTube game. For years and years and years, I didn't take sponsorships. But at the end of the day, if I'm going to make a video anyway about a AI company, I might as well get $3,000 for it. Um, personally, that's how I think about it. But yeah, technical build, absolutely perfect, right? We don't see this very often with uh open source/ just generally Chinese models. So, really, really impressive. Um, yeah. I mean, wow. I would say this looks better than my test yesterday with GBC 5.5. So, I don't know, guys. Let me know what you think in the comments. I don't know why they're getting on. Um, I think this is a really, really good step forward. Um, the SEO is all done. I mean, the SVGs are a bit lacking, but you know, it is what it is. It can't be perfect in every uh every angle, right? I'm going to leave the video there. Thank you so much for watching. If you are watching all the ends of the video, you're an absolute legend. Go and check out New Jane on YouTube. They're absolutely amazing. It will be the song of your summer or the album of your summer. I absolutely promise you. Also, go and check out harbor seo.ai. I have some huge plans for it and you want to lock in that pricing right now. Thanks for watching. I can't remember if I said you're an absolute legend if you watched all the way to the end of the video. So, I'll say it again and I'll see you very, very soon with some more content. Peace out.