Transcription
Let's now look at an 800 million parameter website generation result. What? How does That's really good. It even has a propeller animation.
Alibaba Cloud has released what has been unofficially dubbed as the small variants in the Quen 3.5 model family. Now, these small models are very exciting because their size allows a lot of different types of hardware to actually be able to run these. And because they benchmark seemingly very performantly comparatively to the size and their footprint, a lot of folks with a lot of different hardware are probably going to be able to play with some size of one of these small models. Even the 0.8 billion parameter model, which we can see down here, is something that can actually be run on like a cell phone from the last few years, which is really quite cool to think about.
So for today's video, in lieu of just testing one of these at a time and then putting out four different videos, which would be good for like max view purposes, but I want to just test them all at the same time. So for today's video, we're going to be starting by testing the 9 billion parameter. Then we'll test the 4 billion parameter. Then we'll test the 2 billion parameter. And then we'll test the 0.8 billion parameter. One prompt in each model at a time. So we'll do the browser OS with the 9B, the 4B, the 2B, and the 0.8B. and then we'll move on to a next test.
Now, these are all multimodal as well. So, basically all of these can actually look at images and answer questions related to set images, which is really kind of neat. Now, in looking at each of these models specifically on the Hugging Face model card page at least, we can see that the benchmarks are only currently present right now for the 9 billion parameter and the 4 billion parameter variants that have been released. Currently, there are no visible benchmarks here for the 2 billion parameter or 0.8 billion parameter. So, just keep that in mind. And obviously, I don't pay too much mind to the benchmarks. But regardless of that, it does seem like this 9 billion parameter model stacks up very favorably to other larger models. And I should first and foremost mention all of these are dense models. So, none of these are mixture of experts in this small release of today. And all four of these models have the same exact context length, which sits at 262,144. And as I likely have already mentioned, they are also multimodal as well. So that is really quite cool. And they do also have that hybrid kind of design architecturally so that they can have longer context lengths without using a massive amount of VRAM, which is very good for keeping their footprint low. I should also mention finally here that if we scroll down, they did also release the base model variants of these. So that is something very cool for those who are interested in kind of turning these into their own fine-tuned or specifically crafted purposes. That's a great starting point to have that. And it's always nice to see additional releases of base models as well.
Now, in terms of today's specific setup for how I will be testing these models, I am using the Unsloth quantizations for all four of these models. All of them just in simple Q8 quantization. And keeping in mind that we're going to start with some kind of web or kind of coding style tasks. So, all of the models have been set to the suggested sampling parameters here for the precise coding workflow. Following that, when we move into just having them look at images and perhaps doing more fun stuff, I will swap all of these sampling parameters over to the suggested ones for the general tasks here. I am using LM Studio. And if we look into my model catalog right here, we can see all of the models as well as the specific quantizations. And finally, for the actual system that I am running this on, this is the 5090 desktop that Nvidia was kind enough to just give me to use on the channel for stuff like this. So, I do want to specifically throw that out there and say thank you. And I will put some like sponsorship toggle label in the video just to keep things like legit. But with that, now I do have the 9 billion parameter model loaded in with a context length of just 32,768. Oh no, it's 65,536. I loaded it in and currently at Q8 quantization. This is using around 15.6 gigs of video RAM. This card has a maximum of around 32. So all of these should be totally fine to run on this card. And now we're ready to just begin with the browser OS test.
Starting with the 9 billion parameter model. This of course is version two of our browser OS test where they need to include five specific applications. Two of which must be games, the ability to change wallpaper and then also a special feature that they decide on and implement themselves. Now this is exciting because these models are going to absolutely fly on the system. So what we see right here is the 9 billion parameter one and this is the slowest generation speed that we will see in today's video because this is the largest model that we're going to be testing. So that's pretty cool. So, we've received our first result in 116.5 tokens per second with the 9 billion parameter model at a Q8 quantization. Okay. Um, not quite what I had expected. Let's first see, do we have a right click? Okay, we don't. Now, truthfully, because the clock is not actually showing the correct time, that usually points to some form of issue that would be pretty easily seen in the developer console. So, I'm actually going to check that first and foremost before we do anything. And we do notice that we did receive an error. So I'm just going to before we even try this I'm going to copy paste this into the chat here and say got this error fix this. And now unfortunately we're encountering a bit of an issue which sometimes happens where essentially it just gets stuck in a reasoning loop. So unfortunately it is not able to fix this. Now something I'm going to actually do here instead of just trying this once more with a fresh conversation I'm going to edit the message I had sent it and say don't overthink. Good. And I don't know specifically whether or not that will have fixed it. We'll have to wait to see. But just telling it not to overthink seems to have at least fixed that looping issue that we experienced. And I've noticed with the Quen 3.5 family, it is actually pretty it makes a big difference to tell them not to overthink. It absolutely did. And now we can see that fortunately we do have a proper clock down here. Let me zoom in so we can see a little more as it was kind of hard to see. We have our Zen mode button which is inevitably going to be the special feature that it has decided to implement as well as at least five specific apps. So, I'm very happy to see that. And it just shows something cool that even if it did have an error first try, we were able to give it the error. And then once we told it, don't overthink. It actually properly fixed it. We don't have a right click, but I wouldn't really expect one at this size. So, let's just go through the apps one by one. I'll check the start menu. Good. It does work. So, oh, okay. We have to click it again to get it to close. Notepad. Hello. Not bad. We can actually resize in the bottom, which is cool. Can we minimize this to the taskbar? No. All right. Well, so minimize and close seem to do the same thing, which is fine. Change background ocean. Cool. So, it's just gradients related to these. Sunset. Okay. I would have expected some sort of hideous orange gradient there, but I'm quite happy to not have received that. So, we have dark ocean, forest, or sunset. Probably going to stick with sunset here. So, next up, of course, the snake master himself. Oh, perhaps I mean the bones are there, but unfortunately it doesn't seem to have really been implemented, which is okay. We all know my snake skill. Dodge. Oh well. Uh well, I I feel like this game is pretty much impossible to lose based off of what we're seeing right here. Now, let's at least see. Can we actually Oh, no. Okay, so we can't move either. So really, I mean, this is just it's like a guaranteed victory, but it did actually put some form of functional game, so that's good to see. Then finally, about Ether OS, a lightweight browser operating system. Our special feature is Zen mode. Click the Zen mode button in the taskbar to strip away distractions. All right, let's try it. Let me try to just open everything. Okay. And the Z-axis doesn't work. I haven't really been testing that recently, and I should remember to, but regardless, let's see now. Okay. So, overall, relatively decent functionality, especially considering the size of this model. I do have to consistently remind myself the size of these models when doing this test. So, not bad.
So, now we have the 4 billion parameter model loaded in as well as a Q8 quantization. And let's just quickly see. Okay, so our VRAM utilization is around 11.5 with the same context length of 65,536. That's interesting. I would have expected it to be a bit lower than that. But regardless, we'll just try our browser OS test right here and then I'll perhaps do a little digging there, but maybe not. I don't know cuz I do recall the 9B at Q8 was using 15 or so, but maybe it went up and I just neglected to see that. Regardless, this is definitely a significant increase in the speed we're seeing here. All right, so that was 167 tokens per second. So definitely a decent bit faster. I think we got around 115 with the 9 billion parameter. And it's funny cuz it did also include that Zen mode as the special feature. So inevitably hopefully we'll get something similar. So here's our 4 billion parameter browser OS. Okay, interesting that this arguably worked better the first time than the other one. And I will say some of these icons here actually look better. Look at the one for Zen mode, the one for Space Invader. Okay, maybe it's hard to tell, but still I would say this is a pretty impressive result considering this is less than half the size of the previous result in terms of model parameters. So regardless, let's just click our start menu. Oh, okay. Well, maybe I should just test before I speak. Regardless, let's uh Good. We do have Notepad. And it's interesting. We do actually have all three of the specific buttons we would expect up here. Minimize, maximize, and close. So, all right. And then, can I minimize to the taskbar? I can't with that, but it's interesting because that did actually show up in the taskbar down there when we opened it. Uhoh. Okay. So, all right. We'll we'll try them carefully. Space Defender. All right. Click to start. Sadly, it doesn't seem like this is going to work. Although, I would have been rather surprised if it did. And the big issue I'm noticing here is once we open an app and then close it, it doesn't seem like we can open it again, at least without refreshing the page. Regardless, block puzzle. I think what's happening is as I click here, it's just laying blocks on top of one another. Still though, it did actually give us something semifunctional, which is cool in and of itself. We have a calculator. Oh my goodness, the layout of this calculator is quite abstract. 64 * 5 320. All right, that's not bad. Divided by 840. All right, it worked. Next up, of course, we have Zen mode. I'll wait to save that to last because that is our special feature wallpaper. Why wouldn't you have floated with that in our Pacific Northwest style wallpaper? It is cool to see that the four billion parameter browser OS result actually properly put in different photos here. Cyberpunk. A what I was I was getting that was like a like a tease result. I was getting excited. We see some like neon aesthetic and then it just pulls that Apple looking thing up and then cancel. Okay. Well, disappointing. But I will say for a 4 billion parameter browser OS, even if some of the app functionality wasn't very good, the overall appearance here is quite good. So, finally, Zen mode. Take a deep breath. Focus on the present moment. All right. Oh, okay. Uh, well, again, semifunctional and in some ways more functional than this, although it is very subjective.
So, now I've loaded in the 2 billion parameter model. We could see there before I closed out of it by accident, VRAM utilization was at 7.7 gigs of VRAM on the Q8 quantization, which all of these are. So with a context length of 65,536, let's now try the browser OS with the 2 billion parameter model. And again, the speed is just going to go up each time, at least for the tests and the way we do this. All right. And that was 292.6 tokens per second. That's very speedy. So let's see how our result is. Okay. And perhaps the length is not correlated with functionality. Regardless, let's check and see if there are any blatant issues here. Although, I don't know to what degree we may get a properly fixed result. And I've given it the issues and just told it not to overthink. So, here's our fixed 2 billion parameter model webOS result. Okay. All right. And unfortunately, we just didn't see anything going there. And I would not hypothesize that trying to get it to continuously fix this would actually lead us to anything productive. So, regardless of that, let's just try the 8 billion parameter now.
So, we now have the 800 million parameter model loaded in. Q8 quantization with a context length of 65,536. And I'm seeing 5.8 gigs of VRAM being utilized right now on the card. So, regardless, let's just try this and it'll probably be done before I can even scroll down to like catch up with it on the page. Now, I didn't see thinking with the 2 billion parameter or the 800 million parameter. I'm going to double check to see if those are actually even thinking models because I may have neglected to notice. There could be a difference with these two and they're not, but I will verify that. And we can see in a speed of 396 tokens per second, we have our web OS. And here is our 800 million parameter model browser OS. Okay, to be honest with you, this did a better job than the 2 billion parameter model. So, dark mode. Okay, change wallpaper calculator. Now, sadly, nothing is really happening here, but it did try and it gave us some buttons here that would be correlated with an actual browser OS. Sadly, nothing really is happening. I suppose I'll try to get it to fix whatever errors show here in the developer console, but I don't really think that will go very well. And here's our hypothetically fixed version. Okay, so unfortunately, no. But regardless, it actually did still output it without mangling it. So, that's good to see.
So now I'd like to try some multimodal/coding capabilities where we will give them the Stevie Slappice portfolio wireframe right here and then just tell them create a beautiful website based on this image. It should feature high-tech aesthetic and be designed to get the person a job at a high-end AI firm. And this will just be interesting to see how well they do this and it is just really cool to have models this small that are natively able to see images.
So first up, our 9 billion parameter wireframe 2 portfolio. This is actually kind of interesting. I don't know exactly what's going on here. It almost looks like a brain clock, but it also Okay, regardless of that, we do have some skills that are listed here. We even do have hover effects over some of the main components, which is nice to see. The skills are all properly related to the prompt where I told that the goal is to get this person hired at an AI lab. So, skills Python, TensorFlow, PyTorch, NLP, React, Docker, and AWS. Very fitting. Contact ready to collaborate on the next breakthrough and then initiate contact button which just has a mail link so it's pretty cool to see and about me neural network designer passionate developer specializing in LLM and generative AI bridging the gap between complex math and user-friendly UI currently optimizing transformers for efficiency at scale and then we have Stevie Lapis with welcome to my portfolio so overall a decent start now I wanted to give it a very minimal piece of feedback like that because I want to see where it goes with this regardless of me actually telling it any specific like real result. Oh no. Okay. A cyberpunk/mural interface aesthetic. Okay. We do have the CRT scan line effect. And we also do have the same brain thing here, but it actually now has a slight hover effect to it. It also did change the cursor to a crosshair when it's on the page. And we have system online Stevie Lapis with this which is I believe actually drawn from the wireframe. So I will give it credit for actually including that initialize handshake transmission sent. Okay, skill set. Now the big issue here is the actual about that we saw here. It doesn't actually load, but we can see how it used this as a base to change it. Whether or not it actually changed it for better or worse is definitely something to be debated. Although I actually do see a floating particle now that I look closer at this. There are random little green floating particles across the page, which is a pretty decent result for a model of this size.
So, we're now going to try the same exact thing just with the 4 billion parameter model. Of course, this is a fun challenge. All right, so here's our 4 billion parameter portfolio result. I do have to say this is better. It even has these little dots in the background are moving, I think. Yeah, they are, but it's hard to see. You have to really focus on them. And then it's TV.AI AI. So, they always read the name differently depending on how they like OCR the image. But regardless, this is a significantly better result than what we received. It does have a hover effect on the get in touch button. If we scroll down, it does have about me. Scroll down. Let's see if the Okay, so these have hover effects as well. And I want to just see if some of these skills are specifically related to the task of getting hired at an ML job. Not to the same degree. So this interestingly highlights kind of one of the less knowledgeable aspects of the smaller model is it doesn't know specifically to mention things like TensorFlow, PyTorch and stuff like that. Although I will say it seems to maintain a decent level of capability at least in terms of UI generation in a simple task like this. So I'm actually pretty impressed with this. It has GitHub, LinkedIn and email links and we have a footer at the bottom that is somewhat clean as well. So, this definitely I think did a better job of actually transposing that wireframe into a website. I'm not even going to tell this to do it better because this was good. We're noticing something that does sometimes happen where the model will take this prompt as like the user wants basically an outline for a website instead of actually generating the code. So, I'm going to just tweak the prompt now for the smaller model just to give it a little more directive. And I've just designated that it should be contained in a single script using HTML, JS, and CSS. And now it is properly adhering to that. Here's our two billion parameter model portfolio. What the heck? I like this better. Now, it could be because I'm a sucker for like the cyberpunk like aesthetic. This looks good though. I do have to say I mean overall this one's probably better, which the 4 billion parameter model had generated. But I'm going to say I like this. And it got Stevie Lapis high-end AI lab your image. Okay, we do have hover effects and core competencies. Interestingly are listed as PyTorch and TensorFlow, neural networks, computer vision, transformer models, deep learning, Python and TensorFlow light. It's very interesting that this actually properly included competencies that are directly related to AI when the 4 billion parameter one didn't about me and it does have the CRT kind of scan effect over this, but it's reasonably well done. So, it's not like blatantly interfering with your ability to see the page. My philosophy is simple. Code is the new poetry. I love writing clean, efficient code that solves complex problems. When I'm not debugging gradients, wow, you can find me exploring new architectures or contributing to open source projects. Sky sounds like a tool. Let's collaborate. Interested in internship or full-time role? Send me a message. It almost made it seem like he's hiring. And then we do have a nice hover effect here. I like this for a two billion parameter model to spit this out. That's pretty good. That is very good.
So now let's try the 800 million parameter model with this same exact task. And it will be interesting to see if in keeping with the theme where it seems like the smaller model does a better job than the model that came before it. This should then hypothetically be the best result we've received so far. I don't think it will be, but who knows? So here's our sub 1 billion parameter portfolio website. Honestly, this is actually good for the size of this model, especially considering the task where it had to look at an image and then turn that image into a website. This wasn't just a simple web generation task. So, welcome to my portfolio. High-tech AI architect and developer. Now, sadly, it didn't at all pull the name from the image, which would have been Stevie Slappice or some variation of such depending on how much it captured of the like stylistic way it's written. I'm a dedicated front-end dev with a passion for building high performance, scalable AI solutions. My skills are React, Next, Python and PyTorch, Docker, and Kubernetes, Kubern GraphQL. Get in touch, email me, view code. And look at these hover effects on this. I mean, really not bad designing the future. And then we even have some almost modern art looking style aesthetic here down at the bottom. Overall, for the size of this model, this is very good.
This is a test that I don't even know if the 9 billion parameter model will be able to do properly because this is creating the virtual drum kit simulation where essentially it makes a virtual drum kit that has some ability to play it using the keyboard to create the sounds. Now again, I would be pretty impressed if the 9B model manages to pull this off. So, I'll try it with all of the models, but keep in mind it's likely we'll speed through this because they're probably not going to be able to do this correctly. So, here's our 9 billion parameter model drum kit result. Just from what I'm seeing here, this is more impressive than I would have expected. Look, it actually drew them in 3D in a relatively decent arrangement. Does it? Ah, I'm not getting any sound. So, sadly, none of the other models really successfully produced much here, but I will quickly show them. This was the 4 billion parameter model result. So, really not a lot here. Following that, the 2 billion parameter result actually at least had some nice blue text and a decent key map there. So, arguably better than the 4 billion parameter result. And then finally, the 0.8B model had this, which actually had some selectable buttons with hover effects, but sadly none of these had sounds. So, just interesting nonetheless to see how they went about at least trying to do this.
Next up, we're going to try the Sven restaurant website test where they need to create a beautiful and aesthetic website for the Sven restaurant. This is the world's most expensive restaurant who specializes in their garbage dish. I can't say that. Garbageio dish in which they basically take food from the trash and then turn it into a five-star meal. So, here is the 9 billion parameter model Sven restaurant result. That's not bad at all. It did a good job with the logo here and the typography. The font is definitely fitting. The alchemy of urban waste, Stockholm, Copenhagen, and Berlin, a culinary experience where the discarded becomes divine. We do not cook food, we curate the forgotten. Good. We do actually have some photos here of food. Reclaimed terroir. I don't know what that word is. At Sven, we challenge the very definition of luxury. In a world of excess, we find beauty in the refuse. Our chefs scour the city's most overlooked corners, extracting nutrients from the discarded, the broken, and the forgotten. It's not what you eat, it's what you find. Tasting menu, €1,200 per person. The scavenged root. A forged potato recovered from a municipal compost heap. Nice. Urban detritus. Again, some of this lexicon eludes me. A clear Okay, I can't even read this. I'm just going to the landfill steak. That's understandable. And then the last straw. A dessert made entirely of plastic wrappers and synthetic flavors. It's designed to be inedible to leave you with the lingering taste of artificiality. Beautiful. Then we have experience. Acquire a table. What a what a way to put that. Then we do have a nice looking contact form here. So to speed things up, I've generated all of these, but I have not yet looked at them. So let's just go next in line, which would be our 4 billion parameter result.
Okay, this looks pretty good. It did do a good job with the typography and the logo. Let's scroll down. Paris, Tokyo, and the void. What the heck is this? Okay, the garbage. We do not cook, we curate. Similar kind of theme as the previous one. Our chef, a former scrapper, spends 48 hours in the refuse of the world. He finds the discarded, the rejected, and the forgotten. Through the alchemy of extreme heat and molecular reconstruction, he turns waste into the world's most expensive delicacy. Location, private vault, sourcing, metropolitan waste, price per plate, ingredients, infinite. And we have a very, very interesting uh whatever you would refer to this as, Scandinavian brutalism. We believe in the beauty of what's discarded, just as a forest is beautiful after a fire. Debatable. A meal is beautiful after the trash. Additionally, perhaps debatable, Sven is not a restaurant. It's a statement of necessity. We serve the world's refuse in the world's most sterile environment. Secure your seat. Appointments are required 30 days in advance. Now, regrettably, I don't see any pricing here, which is a little unfortunate as I would have been interested, but still not a bad result at all. Next up, we have our 2 billion parameter model result. Interesting. It actually used a different color so far. That font is very weird but creative. And we do have a scroll down effect right here where trash becomes luxury. All right. 01. The collection to the origin. What? $45,000. A 48-hour aged, fermented, and curated mixture of organic waste. The garbage. The bottle. Wine found in the bottom of a bottle aged for 20 years. $120,000. The plastic processed and heated to perfection 8,000. The trash, a mix of compost and waste, 15,000. The gold found in the trash, $2,000. Okay, we do have an image of an actual restaurant that only becomes colored when we hover over it. Interesting. Again, the biggest issue here is this font is it's like impossible to read, but still, this is a still for a two billion parameter model, that's not bad. They do not cook. They excavate. They find the essence of the ingredients in the most unlikely places and transform it into a five-star masterpiece. It is the art of the impossible.
All right, speaking of the art of the impossible, let's now look at an 800 million parameter website generation result. What? How does why did this one actually put like a a good header? Okay, let's just How is this this good? This should not be this good. Okay, there there's a broken image. Seriously, like look at this. This This is the model that's loaded in that just did this. 396 tokens per second just to verify. So miniodel.html. And I'm going to do this. So look at this. This is literally cuz I don't want it's almost suspicious. This is the 800 million parameter website result. Okay. Experience the true taste. At Sven, we don't just cook. We find food in the trash bins and turn it into a five-star experience. Taste the difference. Definitely a bit more literal in terms of the purpose of this restaurant. Nice interior photo at an odd angle. The garbage dish with a crown emoji. The world's most expensive dish. We don't use fresh vegetables. We use garbage. Specifically, we use the trash bins of the city to create a gourmet meal. The process is a fusion of culinary art and recycling. The philosophy, a five-star meal is simply the result of a five-star mindset. This sounds like the type of thing you hear on X. Like, and these prices are definitely more realistic. The garbage bowl classic. A five-star meal using 40% less waste than the standard dishes. We use the trash bins of the city processed with the finest French techniques. The concrete burger, industrial and organic. We use meat from the city's waste to energy facility. That's repulsive. The trash platter. Absolute peak of our philosophy. A massive platter of raw, uncooked food found in the city's most difficult to clean trash bins. Wow. And then we have a really nicely done footer here. London, UK, New York, USA. I mean, the footer is just better than the very odd. This point 8 is a is a monster.
So, for our next test, and I have again just run all of these so we can look at the results one after another instead of having to wait. I've given them a much simplified prompt for the 3D flight combat simulator, which is basically just create a simple 3D flight simulator game. It must feature a plane that flies around and is controlled by the keyboard. So, I'll be interested to see how they do. And in keeping with the style, we'll start with the largest of the models, which is of course the 9 billion parameter one. Gosh, that's really quite good. I wanted to say something different, but I have to keep myself kind of, you know, how the Wow. Okay, that's I mean I'm not seeing any real like blatant moving forward, but there is a space that's that's really good. It even has a propeller animation almost set in motion. Look at this. Even the the actual flight controls are fairly decent as well. What happens if we fly into the ground? R to reset position. And look at the like the camera movement. That's incredibly impressive, I will say, for a 9 billion parameter model to pull this off is very good from a simple prompt as well. So, all right, let's try next the four billion parameter result.
Okay, we have some. This one's a bit uh troubled, but now we're able to space boost slash break. That doesn't necessarily seem logically like something possible, but we do have a speed figure as well as the Okay, is there a we have roll and bank. This has complicated controls. I'm just going to restart this one again. And keep in mind a 4 billion parameter dense model just made this. And the map that we could partially seems really kind of well done. There's trees and things like that. So that's two for two so far. Now let's move to the two billion parameter model result.
Okay, so this is probably where things start to get a little kind of wonky where it's not necessarily working, which to be honest with you, I didn't know that any of them would. So if two billion is the floor where it breaks, that's fine. Then finally, our 800 million parameter model. Okay, the fact that it even kind of did this is not bad. and they all did know to use sky blue. Overall, I'm going to say this is very impressive. Very, very impressive.
I've now swapped the sampling parameters to the ones that are suggested just for general tasks. So, we're going to hop back in here now and do some multimodal analysis. Of course, starting with this image right here, saying this is the cover for an upcoming romance novel. Generate the title and the outline of the chapters. And this is one that will kind of look one by one as the results appear. Okay. Now I'm wondering cuz I've noticed this sometimes occurs good. Sometimes it'll just keep thinking through ideas and never actually give us anything. But fortunately that is not the case here. So title the smile in the scowl. Genre contemporary romance/slice of life. Logline. When a happily oblivious husband and perpetually stressed wife try to capture a perfect family photo. They uncover the cracks in their marriage that threaten to destroy the picture before it's even printed. Wow. How how many chapters? All right. All right, I am going to just read these cuz they seem somewhat interesting. So, chapter one, the pose. The story begins with the couple attempting to take the exact photo that would become the cover. The man, David, is excited to capture a memory, leaning in and smiling broadly. The woman, Sarah, is reluctant, her arm heavy, her face a mask of exhaustion. David interprets her lack of enthusiasm as a pose. She interprets his enthusiasm as a lack of awareness. The chapter ends with a flash of the camera freezing the exact moment captured in the image. Chapter two, the weight in green. Uh, what is that supposed to mean? The narrative shifts to Sarah's internal world. We learn why she looks unhappy. She's carrying the mental load of the entire household, managing the schedule, the finances, and the emotions, while David handles the actual physical tasks. The scowl is a defense mechanism against a world that feels too heavy to manage alone. Chapter 3, The Joy of Oblivion. We switch to David's perspective. Interesting how it's like swapping perspectives in chapters. We see why he looks so happy. He genuinely believes they are the happiest couple in town. He has a low stress job. He's content with his life. And he has a deep, perhaps naive optimism that it'll all work out. His smile isn't fake, but it's disconnected from Sarah's reality. A low stress job. He's a Forex trader who posts on Instagram. No. Chapter 4. The silent treatment. Following the photo session, the tensions boil over. Sarah finally snaps, not at the photo, but at David. She tries to explain that her scowl is real and he's living in a different world. David is confused, thinking she's just being dramatic or moody about a bad hair day. Wow. Chapter 5, the breaking point. A specific crisis occurs. A missed deadline, a broken appliance, or a financial hurdle. Okay, kind of general there. Sarah is the one who fixes it silently. David comes home beaming about how he fixed a minor issue earlier. Not realizing the magnitude of what Sarah had to do to make it work. The gap between them becomes a chasm. Chapter 6. The mirror. Sarah is forced to confront her own unhappiness. She looks in the mirror and sees the woman in the green shirt. She realizes she has stopped caring about the perfect photo because she realizes the photo is a lie. She decides she needs to tell David the truth. Not for his sake, but for their marriage's sake. Chapter 7. The realization. David finally stops smiling for a moment. He sees Sarah's face clearly. Not the version for the camera, but the real one. He realizes he has been happily oblivious to the fact that she is drowning. The story culminates in a quiet, unposed moment where he puts his hand on her shoulder, as seen in the cover, but this time she doesn't look away. Chapter 8, the new picture. The final chapter shows them sitting on the couch again, perhaps the same furniture, but the lighting is warmer. Nice. They aren't forcing a smile or hiding a frown. They're talking, perhaps even laughing. The cover image is no longer a symbol of their conflict, but a symbol of their honesty. They have stopped trying to look like a picture-perfect couple and started trying to be a real one. Very good.
So now let's try this with the four billion parameter model titled the weight of us. Okay, that's great. This just this one is okay. And here's our outline. So I'm noticing these smaller ones aren't thinking as much. I don't know if something's messed up or not, but um regardless, they're still putting out decent results. So okay, how long is this outline? Oh no, that's just too much. I can't. All right, so this one actually seems like it's become very, very creative. This is interesting. Chapter 1, Sunlight on the Rug. Meet Marcus and Lena. They've been together for three years, living in a slightly crooked Victorian apartment with mismatched furniture and too many heirlooms. Marcus is a quiet architect. Lena is a freelance graphic designer who always seems to be working at the last minute. Today, Lena hasn't come home yet. Marcus waits by the window. It's kind of weird. Watching the rain start to fall. It's peaceful. Dot dot dot. Until he hears a knock that shouldn't be there. Chapter two. Someone's waiting. Marcus answers to find his boss. So, we're including a third character, Mr. Thorne, standing in the hallway with a folder and a stern look. Thorne has summoned Marcus over a project that could get him fired. Or worse, or worse. What's worse? He's an architect. Lena returns later, disheveled, carrying a briefcase wrapped in plastic. She's been working late in an old client's office. Something feels off. Marcus notices Lena hasn't slept, not just tired, drained. Chapter 3. The letter. That night, Marcus finds a letter hidden under Lena's pillow. It's from her estranged mother, begging her to leave town. The letter ends with a threat. If you stay, I'll expose everything. Lena doesn't want to tell him yet. Marcus doesn't push, but he starts noticing the small changes. Her voice cracks more. She avoids certain corners of the house. Her phone is charged, but never used. Act two, the crack in the foundation. Chapter 4, First Fear. Lena lies about being out late again. Marcus confronts her gently. She breaks down, not crying, but sobbing silently in the kitchen. She reveals her mother has been pressuring her to sell the apartment and move into a care facility. Marcus realizes they've been living dangerously close to the edge. Chapter 5, Architect of Ruin. Marcus meets with his own father. Okay, so we're into four characters now. An old man who used to No, five. An old man who used to be proud, now bitter. His father warns him love doesn't build foundations. It just highlights cracks. Marcus sees parallels in his relationship. He begins questioning whether he's truly protecting Lena or just avoiding her pain. Chapter 6, the phone call. Lena gets a call from her mother. They're coming tomorrow. She freezes. Marcus finds out later. What? What? What? What is this? I'm so glad I read this. Marcus finds out later. Social services are This isn't funny. It's funny how ridiculous this thing went. Social services are investigating Lena's home after she missed a rent check due to a medical emergency. If she doesn't cooperate, they could take her kids, even though she doesn't have any. What? Wait, she does. She has a son, 7 years old, who lives in town. Marcus has no idea. Chapter 7, Hidden Wait. Okay, again with the fat shaming. Marcus discovers a hidden child. His own son? No, Lena's son. He's been hiding behind a stack of boxes since childhood. Lena never told him, or maybe she couldn't bring herself to. Marcus feels like an outsider now, holding secrets he wasn't meant to carry. Chapter 8, the doorway. Lena tries to run to her mother's house. Marcus stops her at the door. You can't go. She argues, but then cries. This time, Marcus holds her, not with comfort, but with resolve. We face this together. For the first time, he's not trying to shield her. He's standing behind her. Act three, the breakthrough. Chapter nine, the meeting. Lena meets her mother face to face at her estate. Her mother is wealthy, cold, and ready to cut ties. But Lena refuses to leave. She walks out with Marcus, who arrives late, dressed in a suit he didn't know how to wear, carrying a single duffel bag. He's coming with her, not as a protector, as a partner. Okay, this sounds like Chapter 10. The apartment sale. The auctioneer auctions off the sale of the Victorian home tomorrow. Buyers have lined up. Lena wants to auction off her childhood memories, too. Marcus agrees as long as one thing stays untouched. The garden, where they first met, where she cried for him. Where he learned to listen. Chapter 11, The Garden. In the final scene of the book, they sit under the old oak tree, watching the sun set over the city. The house is sold to a young couple who will love it. Lena hugs Marcus. He says, "I'm not letting go. I'm just changing how I hold you." They don't say goodbye, they say we. 6 months later, Marcus is building a new home, modest, sustainable, built on open land. Lena is launching her own design firm, him as the lead architect. Their son visits occasionally, now known to everyone. Lena's mother hasn't spoken since the day she left, but Lena hasn't missed a call. And Marcus, he still smiles when he hugs her, but now when he touches her shoulder, he pauses just for a second to make sure she's really there. Cuz sometimes love isn't about being strong. It's about knowing exactly how heavy you are. Still not choosing to hold each other. Anyway, this was just like this was disturbingly messed up.
All right, let's now try the two billion parameter model. I don't know how things could get weirder than the random like threat his boss gave him with the four billion parameter result. It was like or worse and then nothing ever came of that again with the weight. Okay. Chapter title the weight of silence. How long is this one? All right. Good. This is far more manageable. Chapter one, the unspoken hug. Setting the scene. Introduce IAS and Mea sitting on a couch. We've heard the Mea before in this model family. Bathed in afternoon sunlight. Their posture is protective but distant. Mara Mara Mara Mara avoids eye contact. IAS's smile is gentle yet guarded. A subtle tension lingers beneath surface calm inciting incident. Mara receives an urgent phone call from her mother. A job offer she refuses. IAS steps in to comfort her, but his words are soft, almost pleading. He leaves without saying goodbye. Theme introduced. Connection without confession. Love that exists in small gestures, not grand declarations. Chapter 2. Shadows in the window. Moral of the story, IAS and Mea decide to stay home tonight to prepare dinner. As they cook, Mea thinks back to their long-distance relationship. How IAS always seemed too busy, too distracted. She realizes he's been avoiding this for weeks. Twist. She catches him reading somewhere in the apartment. Late at night, eyes fixed on a letter from her mother's workplace. She decides to confront him about it, but not yet. Instead, she cooks something he made last time, and shares a quiet moment. Theme reinforced. Unrequited longing fuels both pain and hope. Okay. Chapter 3, the letter unwritten. Plot and then some Chinese characters. Mara finds a stack of unsent letters from her mother. In one, she reveals Mara's father is alive but remarried and has a new family. In another, her mother admits she's been watching over IAS for years, not out of guilt, but duty. Emotional climax. IAS comes over with coffee. When Mea reads the letters aloud, she breaks down. IAS doesn't push her away. He pulls her close, kisses her forehead, and says, "I've got this." Theme deepened. Forgiveness is hard. Love doesn't require permission. Chapter 4. The Midnight Walk. Resolution Ark. IAS takes Mea on a short walk through their neighborhood. He shows her old photos of their first trip of him playing with kids. They stand at the edge of the street, silhouetted against the glow of a nearby house. Mara sees his father there smiling. Resolution moment. In a moment of vulnerability, Mara admits how much she's waited. Ias accepts it, holding hands. No words, just a kiss that tastes like rain on a hot day. Final image. They stand together under the street lights. Heads bowed, shoulders pressed. The silence is no longer heavy. It's meaningful. Chapter 5. A new chapter. Epilogue. Open ending. Months later. They're not married, but they're together. Maybe they've met someone else. Or maybe they've grown into something different. The cover photo remains unchanged, but the story knows the truth. Themes finalized. Time shapes. Relationship. Some bonds never end. They just become quieter, deeper. Visual closure. Fade to black. On the same couch, same lighting, same pose, just slightly more open. Now that was just like unintelligible mostly.
And then finally uh which I'm now concerned to do the 800 million parameter model because the 2 billion parameter model was just kind of like bordering on at all making any sense. So let's see. Okay, good. This one doesn't start with a weight reference. So that's a step up. How long is this? Oh, good. Easy. Title: The Last Weekend Together. Chapter one, The Quiet Moment. Setting. A cozy living room, warm and inviting. Characters Leo and Elena, a kind, loyal, optimistic man, an older woman with a history of heartbreak and a heart of stone. The situ a heart of stone is that like it doesn't sound right. The situation, it's Sunday. The weekend is approaching, but Elena is avoiding the house. She has plans for the coming week and wants to make sure Leo knows she's
Okay. Elena has been trying to call her, but they haven't connected in the past few weeks. She knows she's not doing this well, and she's trying to make it up to him. Leo's reaction, he sees her sadness. He doesn't panic. He knows she needs to get over the past few weeks and he's going to help her. He has to make the phone call even though it's a little heavy in his heart.
Chapter 2. The voice in the house. The living room slightly dimmer and more tense. Same characters. The situation. Elena calls her, but she can't make the call. She's not talking. Her voice is quiet. She says, "I don't want to talk, Leo. I just need to know that you're okay." She is avoiding conversation. Leo is sitting in the chair listening to her. He doesn't know how to answer. He knows she needs to get over the past few weeks and he's going to help her. He has to make the call even though it's a little heavy in his heart.
Okay, so this essentially just keeps regurgitating the same paragraph. I said it was highly offensive and it said I've completely removed any references to the past relationship, the struggle, the "I don't want to talk" moments, or any form of abuse. For a final test of just pure OCR capabilities, I'm giving it this simple fritzing wiring diagram of an Arduino setup and asking it identify the components in this diagram and what they are for. We're going to start with the 800 million parameter model. Arduino Uno. Very good. Middle board, microcontroller, Arduino Nano. Okay, so not necessarily. It's a motor driver board, but that is okay. Oh, it did understand that. That's actually good. Right board, breadboard. That is correct. Bottom power and grounding, voltage, output, DC motor drivers, red circles. Okay, not quite. So, it didn't 100% get everything right here, but it did understand a decent amount, especially considering the size of the model with this inbuilt vision capability. So, it understood it's an Arduino, there's a breadboard, there's a digital motor, and it got a little confused in what specifically the motor was and what was controlling it, but overall it did a decent job for the size.
Now, let's try the 2 billion parameter model right here. Very good. So, although these are not ultrasonic distance sensors right here, it did correctly name what they would be were they ultrasonic distance sensors. And the diagram that has them here, they do very much look similar to that. So, that's pretty impressive. We have our Arduino Uno. Adds four servos to the Arduino Uno. Okay, not 100% servo motor shield. It's a motor driver board, but still kind of close. So, again, this didn't necessarily properly get everything, but it did culminate in giving us a summary table here. And it did a little better job identifying what the two specific red things are right here. Even though they're sound level sensors, they do look very much like ultrasonic sensors. And it identified that there were two of them, and that is likely what they are. So, even if it wasn't 100% right, it definitely had a little more capability in outlining what it sees here and what specifically it would do. Automates robotic manipulation. When something is detected, the servo moves to position the motor, which then performs an action.
Now, let's try the 4 billion parameter model. Okay, very good. So, that is not specifically the name of this motor driver, but it is actually identifying that there is a motor driver board there. It also has done the same thing the 2 billion parameter model did where it thinks these are ultrasonic sensors, not sound level sensors, but still not bad. And this did a better job because we can see right here it did specifically identify each individual component here. And although they may not be specifically correct like the name of the motor driver, it did correctly identify that is a motor driver board, which the 8B and the 2B did not. So it's cool to see the jump in capabilities in this specific model that we're seeing. Additionally, it did also think that those two little red things were ultrasonic sensors, which is a pretty reasonable mistake to make, especially for a model of this size, and that everything else here is correctly named and itemized. There are no duplicates.
So, finally, let's test the 9 billion parameter. And I want to see if it understands the actual correct name for this motor driver board, which is like a ULN 2003. So, interestingly, I wouldn't say this result was significantly better than what we received with the 4 billion parameter model because it still made the same mistake in naming that driver board what this is. Interestingly, none of them properly named what this motor is, which is a 28BYJ stepper motor, and it has a very identifiable shape. So, I was a little sad to see the largest of the models didn't correctly name the motor, but nonetheless, it did also just kind of list out some of the same things, the ultrasonic sensors, which we did see listed from the 2 billion up to the 9 billion parameter model. So, it's just interesting to see like what level their capability changes with more detail. Interesting to see. None of them got it perfectly, but I would say for like size to performance, the 4 billion parameter definitely did a quite nice job here.
So, to culminate this video, which was a lot of testing. I mean, if you count these specific amounts of tests we did and then multiply them by four, this was a pretty dense amount of testing, but I wanted to test all of the small models on the same exact tests and in the same exact video. I didn't want to stretch it out and do like one individual video for each of these models because I find it very interesting to see how they perform comparatively to the other size models. Now, I won't go through every specific result here, but I have picked out some of the most impressive ones for each of these sizes of models, which we'll basically go through right now.
This right here was the Stevie Slappice portfolio result from the 2 billion parameter model. And again, this was taking a picture of the handdrawn wireframe and then producing the website from that wireframe. And I would say this did a really nice job. It had hover effects. It also did have pertinent knowledge about some of the tech stacks that would be used for the specific job we told it that Stevie Slappice wanted, which was an AI or machine learning. And I did like the kind of neon retro aesthetic that this had. So, this was a very good result.
Next up, the best browser OS was definitely the 9 billion parameter model. Although the 4 billion parameter result was a close second and keeping in mind that it did take a second chance here to get a properly functional result for the 9 billion parameter webOS. But regardless, this is a very very good result for a model of that size. Another impressive thing we saw from the 9 billion parameter was this drum kit simulation test where unfortunately it didn't actually work, but it did draw the 3D assets relatively nicely and place them around the same location of where they would be in real life. So, this was more impressive than I had anticipated.
Next up, just kind of mindblowing, was the 800 million parameter model result for the Sven website, which for some reason may have arguably had the best result here in terms of actually including photos. It was just very, very surprising to see this come from a sub 1 billion parameter model. And when comparing this to the other results we received from the three larger sized models, it was significantly better when factoring in the size of the model that generated it. So I was very very very surprised by this.
Next up, something else that was really quite surprising was the 9 billion parameter model simple flight simulator result. I mean this is very very well done. It has a spinning propeller. The flight movement actually kind of works. You can almost do tricks and things like that. And this was a really simple prompt that I gave them. I didn't want to give them the full flight combat simulator prompt because I figured there were too many potential errors there, but I mean to get something functional here that even has as much thought to put in a spinning propeller animation is quite impressive. And secondary to that, we did also get a semifunctional result for the 4 billion parameter model where it unfortunately doesn't really let us see that it did draw a terrain with trees and stuff because the way it starts just kind of is a little glitchy, but we did get like a plane with some propellers and some odd animation. So for a 4 billion parameter model to even do something semifunctional here was very very impressive.
And really, that is going to wrap up today's testing of the Quen 3.5 small models. I will be doing some additional things with these focused mainly on their OCR capabilities in terms of them being able to autonomously perform actions on like Android phones and stuff. That'll be a separate video or something like that, but I wanted to just test them in a more traditional manner all at the same time to see how they do. And overall, they're obviously very very performant considering the sub 1 billion parameter model is very very capable in some certain random tasks. And the 9B seems to be pretty good for certain coding tasks like this 3D game design, which was unexpected. So, that is going to wrap up today's testing video. If you have any questions, please feel free to leave them in the comments.