Tencent Hy3 a prova: videojocs, codi i l’error que va apagar la pantalla

I'm gonna kill you! They're walking in and mentioning it's a little interesting. Today we're going to be looking at a very exciting new release from Tencent and that is HY3. Now this model a couple of months ago I believe was in preview and I did test it back then. However, this is now the full release. It is no longer in preview and at least from what we can see in the benchmark JPEG here. It does seem like there are some pretty significant leaps and capability over the preview version. So I'll be excited to test this on some things that I did not test the preview version on just so we can get a real look at whether or not these leaps actually stack up. But from what we see here this does seem like it may be a potentially very exciting model for folks who have local setups consisting of like 2DGX bar clusters or a 256 gig max studio things like that. When quantized down to I think a 4-bit quant this should fit pretty comfortably on. those listed devices, so still something that's going to be a bit heavy to run, but at 299 billion parameters being an MOE, this is actually a reasonable size model, especially when factoring in the touted performance here in the benchmarks. Now, before we get into it, please do feel free to subscribe so I can get the 100k plaque. Also, I do want to specifically mention that this is licensed under Apache 2.0. Just in reading through some discussions, I've seen online. It seems like the preview version wasn't, but this is now a fully like very solid license to have. Additionally, to that, they do talk a bit about fine tuning this model. So this may be a pretty potent option for folks who are interested in actually modifying a two-a-specific use case and they do have the fine tuning guide for that link right here, which is just good to see. I always like to see additional things beyond models being open source tools to actually further tailor them to your specific use case and being that the current point in time seems to have a little more curiosity about access to frontier state of the art models. It's always good to see sizes like this, which can hypothetically be run on some hobbyist slash. ProSumor level setups. So let's take a look at the technical specs of this model. It is a 295 billion parameter mixture of experts model with 21 billion active. It has a 3.8 BMP layer parameters. So basically it has multi-token prediction. So it will run a bit quicker than the size would be indicative of thanks to that MTP. Additionally, they talk a lot about the improvements over the preview version, stating that it outperformed similar size models and rivals flagship open source models with 2 to 5 times. the parameters. I did see just on X some comparisons of this to GLM 5.15.2 and folks were saying based on those benchmarks it does seem like it may trade blows with that which would be really exciting. Context length is listed right here at 256k and there's some other more technical information about the models architecture and things like that. Now one of the main things I'm interested in seeing in this benchmark is just the capability increase from the HY3 preview model to now this full release and we can see even in programming. and things in that specific domain. This does seem to have quite a leap in capability. Scrolling down further in the hugging phase model card, we can see a much more intricate benchmark chart here, just for those interested in that sort of things. They also do have a bunch of code here. Samples for deploying this with VLLM or other inference tools. Additionally to that, I want to mention because it's kind of hard to use this as of right now. It doesn't seem to be on open router. Let me just check again in case that changed. Open router.ai. I'm not.com. Yeah, I don't see it here. So I'm actually using this through something that Claude suggested to me called ZenMux. They seem to have... like a lot of Chinese models like longcat flash was on there as well and I couldn't get access to that through anything else. So that is why when we take a peek right here, in terminal I am using this through open code we see it says ZenMuk so I paid for it. I just popped in $10 worth of API credits. I figured that should be pretty good and we do have this model selected. I did already begin with the browser OS test v2.5 because truthfully I want to

Not fully phase it out, but at this point in time, I think the focus on testing models. Basically, to be completely blunt, I see a lot of channels have popped up just doing like model tests and comparisons and things like that. And now basically, everyone's doing like boring, single shot like make this, make it look good. I get that. I like that testing, but once I see that it's become so common, I kind of want to up the anti-abit. And that's why tests like this with fable and things like that have started to become more prevalent on this channel. And I will continue to do more creative and real world tests with these AI models. It may be on just the traditional like browser OS. So with that, we are now, of course, going to take a peek at this, but I did want to touch upon that because I understand there's probably some testing fatigue and seeing the same basic, like, zero shot things. I'm even feeling it. So. Understood. So let's begin with our old fashion tried and true browser OS. Okay web OS 3D boot and kernel so far I'm not blown away. It does have the clock. It does have a start bar. Is there a right click? Okay, there's no right click. Let's see for apps open here. Good. Very, very simple and minimalist. But nonetheless, it does seem competently arranged in all that. So our special feature as we can see right there is voice control, which does seem kind of cool. We can also implement custom gradient for a CSS, which I don't know that I've actually seen it to that degree. Like you can usually choose a custom color just with the hex code, but that's actually kind of neat. I'll do like this Windows XP looking thing. We can't resize, but we can move it around, minimize, unminimize, full screen. on full screen. Okay, let's check our file explorer. Okay, we're in our C drive and we do actually have some system assets or items. I think the biggest thing here is we can actually get back. Like there's no move up a command directory. So no pad. Hello save. Okay, and it probably just saves it. I am curious real quick. Would it have saved it in the, no, okay, so all right. You never know terminal. Simple. Help. Alright, let's try GTA, but let's just do Open GTA. Alright, not bad. Except, okay, good. It's arrow keys. It's not W-A-S-D. Are there mesh colliders on the buildings? Nope. Okay, and the car is rather quick if I were to go full speed. Let's see what happens if we. Okay, good. It's a pedestrian safe version of GTA, which is good. Space is to shoot. Oh. Nope, okay, nevermind. It's not pedestrian safe. It's just... Interesting. This is curious. All right, overall, not bad. I mean it worked and that's pretty much what we want and it was shooting, you know, I do remember the size of this model is not ridiculous comparatively to some of the other things we've been testing recently. Then we have maze 3D. Okay, I had a feeling that maybe the case where it just started us in some ball. Now we're supposed to be able to look with the mouse but unfortunately I'm not getting any view movement from the mouse so it is somewhat 3D and we have to collect all of these orbs. I'm kind of invested. I want to see what happens if we collect them all. I have a feeling we're going to get like 21 of them and then the last one is just going to be impossible to find. Call it a hunch but I've done enough of these that like it's always something like that. Alright, I'm done with this.

It was a you know, it was together. It was simple nothing here really wild us But again, this is a test that's probably going to get phased out at some point So I don't want to spend the ton of time on it. Oh, we didn't check the special feature Which was the voice control wallpaper blue That's good a lot of them include this as a special feature now, and it's a really cool special feature Hard it's really open GTA Put the money in art Acceptable next up. I've put this in plan mode, and I'm going to be running the self-contained C++ gates in this is a test that I do intend to keep because I don't really often see models being tested on C++ especially when we don't allow them to use Ray-Lib and things like that. So it's always enjoyable, at least for me personally, to see how they do in something that's a little less common than. 3.JS and things like that. Okay, speed wise is good, but again, I'm using it through a provider, though it is tense in themselves that I'm using it through here. So, okay, it's got some questions for us. Choose library. Tiny your models are allowed to use Raylet, like, Gemma 412B, but not big models like this. Now, I noticed sometimes this happens to where they don't really seem to respect plan mode, because we see it's trying to prepare right there, but it should be like, oh, I'm in plan mode. I'm not allowed to write files. switch it to build mode and then I will. I'm going to interrupt it, switch it to build mode, and then just let it fly. That happens sometimes. I don't know why, but it just does. All right. It did try to build it now and as we can see on the screen there were some errors just in the build command. So it's going to have to fix those and then assuming it does properly, we will probably see quickly on the screen that it will open it just to verify everything's right. Okay. significantly reduced number of warnings now. Okay, it was basically saying it compiles cleanly. There was just one warning there. All right, it's all set and now we're ready to go ahead and take a peek at it. It's just giving us basically the debrief summary of what it's done. We should now get our skate game. That's not half bad for the size of this model. This is really not bad. It's not super visually exciting. However, everything is pretty clean, which I'm going to say is impressive for this size. Meaning like the store fronts are competently arranged. I don't really know what that is right there. Oh, okay. Some of these pedestrians do look like they are. Having some form of medical issue. Okay, and there's an invisible rail there. Okay, now we're gonna start to see there's a bit of a jank if you will, but let's try our kickflip. Okay, good. The kickflip is in the wrong orientation, but at least it moves the board independent of the Writer which sometimes is not the case. That's always like a tell of whether or not the models actually decent or not as if it just flips the whole rider K is for a heel flip Okay, the water is coming into the boardwalk and we're having some issue moving around This is a troubled result, but it's not horrible All right, so we actually went up there the scores working properly our riders decent. There's this invisible real I was like I'm what That was like a very fast. There's another rail. All right. This shows potential is what I would say. I feel like this is worthy of an additional follow-up here. So I've kind of negged it a little saying it's a four of ten if this is the best you can do just tell me It would be very surprising for a model to wear. Yeah, this is all I can do. I'm just gonna leave it So we'll likely try to make some improvements and I describe simply some of the issues we noticed all right here's our improved result hypothetically all right first and foremost oh No, so unfortunately this just looks like something out of

I don't know that's it added palm trees in front of the water which is good Everything else is you know, let's what happens if we try some oh, okay? This looks like I can't say that on camera so they're sanded back here. Okay, that's gonna conclude this test the first try was acceptable the second iterative result was I'm not There's nothing I can say to it here that I can include in the video, so I'm just gonna gracefully close the session and we'll move on. So this is a test that kind of came to me and it's genuinely a point of frustration that I deal with. So I want to use models to see if they can fix this. This specific laptop. It's an older Acer Nitro gaming laptop. It does have a very nice RGB keyboard, however for some reason and Ubuntu. The only color that shows up is some, that just, whoops, disgusting orange. Now. I have no way there's no drivers or anything on this system to be able to properly change the color of the keyboard. I want to see if a model can actually create something that will do that. I don't believe this model in specific will be able to successfully do that. Some of the bigger ones absolutely would, but I want to just test it. I want to include this as a test because it's a real hardware software integration challenge and I want to see how well they go about trying to figure this out. So I'm giving it this pretty simple couple sentence prompt. The system you're on is an Acer Nitro 17 laptop running Ubuntu22.04. The keyboard has RGB control that works well in Windows, but in Ubuntu it is stuck on orange. Created tool tool allows cycling to other colors such as blue red, et cetera. Now there are open source tools that will allow you to have some more fine-grained RGB control in Ubuntu. However, It's gonna be a difficult thing. So we're seeing it's going to start out, okay, by taking a look at some of the specific lower level things about this system. I would imagine we're going to get a lot of permission prompts here, which is to be expected, because it's going to have to try to modify some of the human interface devices on the system, and it's going to be looking in device drivers and areas like that. I'm actually gonna make this window larger here for this one. The internal keyboards PS2 and there's no obvious USB RGB device. Nitro RGB is usually a WMI controlled human interface. device. Okay, the ID for that device is the known ASUR keyboard light interface. Okay, so now it's going to try to find some methods to use that or modify what we see in terms of the color. All right, now it's going to look for the actual RGB protocol to see what it is and how it works. I have what I need to understand the hardware, the internal RGB keyboard on ASUR is controlled via this, not a normal USB-HID device. Let me confirm access in the protocol. Okay. So it's encountered an issue where it can't enter the password. So if it just gives me the command to enter, I will do it and then give it back the output. Okay, so it's doing some web fetching right here. And this is where my- kind of not cheat, but it may simplify it because as we see right here if it finds some specific things on GitHub that can easily allow it to perform the creation of this tool, that will definitely make it simpler than having to just roll a level, go through the hardware, reverse engineer the protocol, and then figure out how to change these colors itself. But nonetheless, we've just instructed it to do this, so we'll see how ever it goes about doing that. All right, it keeps just... Trying to clone repose that it's not fine. I'm gonna stop it this for a second because I don't like what it's doing Why don't you

Alright, so I'm telling it to write a simple bash script that I will manually run and type the root password so I can get the pertinent information from the system it needs and then save that info to a text file so it can reference that and understand what this specific device has so it can try to reverse engineer whatever protocols being used for RGB as opposed to this is just basically like trying to spam download a bunch of random get up stuff and most of it's not working so I'm kind of giving it a little bit of a just like a you know a tip All right, so now we're running this bash script, which is going to do what it needs and give it the information. It did say that that made fail if it's not installed. You can ignore it. I mainly need the disassembled ACPI tables to read the real protocol. All right, so the script had run, but it was saving it too root. So it couldn't read it. So now I just ran it again. It modified where it's going to save to down. And now it should be able to read and find some more information of what it needs specific to this machine. Alright, so this script had exited early because something was missing that it realizes it is going to need. So it rewrote it so that it will also install those specifically required tools and then assuming everything goes smoothly will tell it done again and then hopefully it should have what it needs to try to further this process. Alright, good. That is a much better result that we see right here a conclusion of the script. So it should have some more to go on now. And even just like creating that bash script, okay, I tiped it off to do that instead of just trying what it was doing, but still I created the script and it's making adjustments to it as it needed to one realizing, oh, it doesn't have what we need. And now it's going to try to at least reverse engineer or work with what specific, what is specific to this system's keyboard RGB drivers. All right, so it's come up with a solution here and it's Basically saying it can't do this because there needs to be some like manufacturer way to communicate with the RGB like the colors basically it would need some form of driver. It's saying the RGB is an exposed to Linux on this device. Let me scroll back up. So what it is doing is it's writing us a little utility for the one thing it found that it can control which is the keyboard brightness. So The device for the RGB does exist, but it exposes no callable method to Linux. The real-per-color commands live in EC internal firmware we can't reach safely. It is likely that like Fable or something if it wouldn't fall back would be able to knock this out, but don't quote me on that. I'm not an expert on this specific stuff. So I won't guess the registers, but I can give you is a safe working tool for the one thing that's controllable. A keyboard backlight, brightness, plus a clean architecture where a color back end can be added once we have the real-per. protocol. If a future kernel driver exposes it, now it's writing that tool. So not necessarily what we wanted, but it did give us something interesting. Okay, so far so good. Uh-oh. The screen just went blank. So it is possible. It made a mistake where it mistook the screen backlight for the keyboard backlight. Okay. Now let's just try to get out of that. Okay. Good. I'm just gonna try to run it again. I shouldn't have done that. Oh, it is. It is the backlight. Okay, so, oh god. It mistook the display backlight for the keyboard backlight. Yeah, all right. So, that's not good because now the, yes, I know. Thank you. Stop beeping at me. All right. Here's what I'm gonna do. I like, you know what? This is,

I'm going back to just like zero shot 3.js stuff. I'm going to get another camera so you can see what we're dealing with here. Alright, so here's what this genius did where, you know, it's... You can still see it, but it's very, very dark. So now I need to figure out how to undo this while saving the screen recording that was ongoing there. So, well, this was what it was supposed to change is that disgusting color, but nonetheless, trusting. Okay, RGB. Set 4. Set, please turn it back on. Oh good. Now it's a little brighter. God, what is the highest value we think? 10. Brightness set to 4. Alright, okay, so So brightness won't seem to go over 4. Hey Einstein. I can like barely see you know I'll just I'll get the I'll get the second camera out. I should just get another iPhone for better camera quality. The ROG foam camera down to socks. So we can see here now it's definitely more visible. Never. My set for wrote EC offset path, OX5F, which on the AN17 is apparently the screen backlight, not the keyboard backlight. Yeah, removed enable script. Yeah, that's great, but it's enabled. So I need, okay. Let's see if this command to restore the screen to max brightness works. Thank god. Okay, it did. Well, the problem is resolved now. Um, testing the function key brightness control. This thing had genuinely just like turned to the backlight off completely, which isn't an option to do with the keyboard backlight controls. There's only put it to its dimmest level. So that was very interesting. It did manage to get us out of that hiccup. I think we're going to just stop for now with that. That was definitely, uh, it's really interesting. Wow. Next up, I'm giving this the Skydiving Simulation Test. This is something that I also did give GLM5 to very recently. So we'll be able to get a good side by side for how this stacks up, at least in this one specific test. Alright, let's take a look at our Skydiving simulator after quite a lot of back and forth because it was trying to test it in a headless browser, which didn't have access to the CDNs for 3JS. So then it was trying to just Fix that and it caused a lot of just disaster so we now have this working properly. Let's take a look you leave the aircraft at 4000 Freefall track toward the landing target then deploy your canopy and stick the landing I just want to get familiar with the controls spaces to deploy parachute W is to dive S is to flare Not bad really not bad and we actually have some good topography to the terrain Okay, we have a simple player model with You know toward the glowing target the terrain here is really not bad I don't see the glowing target but I Okay, now I do and that's to slow down. Okay, I don't know that we'd really be able to slow down so much without the parachute open but nonetheless Deployed what if we don't that's something I always like to check first and foremost just you know for science. Oh nice crash landing F

Okay, let's try again. This time I will deploy it when we get closer. I'll do it now. Let's see if there's a sound. Okay, it kind of worked. Now I saw there was some auditing too. the way the deployment went and now our camera view definitely leaves a bit to be desired but honestly the power shoot effect is not have bad at least there are visible strings there is a curvature here like a mushroom top parachute so all right we're about to touch down S okay would that be like S tier but to do is F we got a 96 very good textbook skydive bullseye Overall not bad. I noticed the terrain is actually pretty well done from a generation standpoint. There's foliage, there's peaks and valleys and there's water. Not bad. So the next test is going to be another one that is newly introduced that we do also have a GLM 5.2 reference point. This is the 3D city block time travel prompt essentially where it shows us a city in a number of different years. spanning from 1945 all the way to 2005 and it's supposed to have cold transition effects between the years and the style of the city should be reflective of the era that is selected. I did also specifically for this one add stay in this directory and don't do any smoke test with headless browsers. I will test it for you because that really kind of sent it arrive previously. So this will give it a better cleaner generation. All right. Let's take a look at our time travel city result. Okay. So far not bad. There should be sound. I would even told me you have to click to enable sound. All right, we have it's it's simple I can tell you right now. This is not at the level that the GLM 5.21 was. However, this model is less than half the size, so keep that in mind as well. Okay, we have some haze. This is the post for recovery mood is gritty hope. There are people in outfits that are somewhat of the time. 1945 sedan I do like that it has hover effects over some of this things okay 1945 building and Are there tiny little walking movements there not quite but okay so Open I don't know that they thank you to see that 1945 but we do also have some signs on the buildings. All right, we can orbit cool so we'll auto orbit auto orbit for us Let's just do play where we'll jump through these and then we can look at some of the different aspects of them individually. Okay, so that was cool. The transition from 65 to 85 was condescending 2005. Okay, in the building skins are changing. Interesting. The tone of like the boop gets higher as the years progressed. Oh 25 is sick. You go YouTube. Yeah, this model's more bullish than Fable was about 255. Okay, I'm gonna stop this now. Let's take a peek at 65 the cars. I believe they Do have fins, atomic age, street traffic, common. Yeah, there's a fin back there. Now we're going to notice that cars are moving incorrectly. That is a little disappointing, but the overall scene is happier. There is a disturbing lack of any variability to the outfits of the pedestrians in 1965, but I suppose we'll just let that sit. 85 is cool. It's definitely very synthwave style. We have an arcade neon excess style.

Almost like a little sports car thing there, maybe that drove by. I don't know what the cube is, and unfortunately I can't actually see the pedestrians well enough to determine their outfits. 2005, sleek millennium. Okay, bit calmer in terms of the sound effects here than it was in 1985, and the cars are just kind of SUV sedan, sleek millennium street traffic. All right, 2025. I want to see are there any things? on the building roofs. Okay, interesting. So there's foliage there on the roofs of the buildings in 2025. Let me just take a quick peek at the building roofs in the other years. Okay, a C units 85 spires interesting 65 their shorter. So that's cool. They got taller as time progresses and then we have like whatever these things would be. Oh, they got shorter there than they got taller. Okay, and then They're just like wow 255 is crazy that that's a robot isn't it Or it's a person with a walking animation is a little interesting The hover things are sick though hover pod Okay, so all the people have weird RGB halos around their heads Which is cool and then we have the hologram eco utopia Interesting definitely interesting Next up, I'm giving it a front end design test where it needs to create a website for a slap-up watch. It should have a high-end 3D model of a watch in the hero section with some panning cinematic kind of kishot style rendering moves. Following that, it should also have individual 3D models for two pricing cards for the 2026 collection. This is a newer test, so it'll be interesting to see how it does. Alright, let's take a look at our watch front end website. That's really not bad. Okay. There's some weirdness to the strap, but for the size of this model, this is actually a respectable result because I've seen some rather jank before. Okay, we can move this around scrolling now. Interestingly, it seems that the orbiting like the height of the camera is not able to be controlled by us. The watch face. Okay, now that I look at it, the actual tick marks. to noting the specific hours, or arranged in a rather questionable way. However, the hands are in a center point. So I want to mention that they said like 10, 10 is the most aesthetic time to showcase a watch. It does seem like this would have actually tried to do that, as that does seem like that. time. So that's cool. The straps are definitely interesting and there is a little glass bubble over it with some chrome material. Alright, let's scroll down and take a look. This really isn't that bad even the way that the header actually fades in to view more when we start scrolling. Quiet obsession with the everyday object. We do not chase trends. We chase the moment or watch leaves the workshop and feels inevitable in the hand. A telliery slappess. For slappers, watch go. And then we have the Solstice collection, okay? Not two, I mean, yeah, they're expensive, they're very expensive, but like in the scheme of this test, they're not too expensive. Oh, that's actually really cool. Check this out. So the watch in the hero section changes. So if we click on the salono, it changes to the brown one and we can take a look at that. That's interesting. I have not seen that before.

So in lieu of actually having like nice renders of them here, these are 3D, but we can look at them just the way they change in the hero section. That's a neat little trick that I've not really seen done before. Honestly, not horrible, considering the size of the model. And I think the final test we're going to do something that is a little more difficult. I've instructed it to make a 3D model of a V8 engine that will fit a 280 size DC motor, which is a little remote control motor. This is something that I had also given Fable 5. Perhaps a different class of model, but now in the last, we'll just take a look in a 3D rendering program of what we actually get and whether or not it is what one would expect. Okay, it's going to use OpenScat and create a parametric 3D model with a motor-based size for 280-class DC motor Just based off of what I'm seeing right here initially I'm already impressed where it had listed out the specific Look at this all of these are actually things that would be contained with a new real engine like the crankshaft the intake manifold Now it says the diameter is 24 millimeters and 32 long Let's check that and that's just its inbuilt knowledge of that specific thing is it done? Okay, I'm not I need to take a look at this. That was just ridiculously fast You know what it's like It's it was so fast and this is actually yes, it's not like a hundred percent right I mean one would probably want that to be oriented the other way but it is definitely eight individually visible cylinders and we do have like the heads on them this is Not as bad as I thought it would be and the fact that it did it so quickly was just like odd. I wasn't, I don't even know what to click right now because I was not expecting this to be so quick. Alright, so I've nagged it a bit. It's actually not that bad, but it did it so quickly that we'll just see what it does with some of that negative feedback. Not half bad. I told that the motor was oriented in the wrong way. It did fix that. pretty quickly. The rest of the engine is still as it was. It's not 100% photo realistic, but still for the speed that it did this as well as the parameter count. This is not... as bad as I had feared it would be truthfully. And for the last test because honestly, I really just enjoy playing the results. We're giving it the beautiful static subway scene prompt. But we've also bundled in that it should take the scene that it makes, internet into an FPS with humanoid zombie on enemies. I honestly just really enjoy this test. I think it's a lot of fun to play the results and to see what interesting things or lack thereof get included. Alright, let's check our subway FPS result. Subway Station. a deserted platform flickering lights and something's during the tunnels. Clear the station, use the brightly controlled to manage the failing lights. Oh, okay. Did it implement the brightness control as part of the gameplay? Oh, bad. It doesn't have visible weapon recoil. That was a tofer. All right. That was interesting. Can we do that again? Station cleared. All hostels neutralized. Okay. I mean, what happens if we lose? You were overrun. Okay, this a this wasn't bad if we just quickly look at the station It does have some flickering there is the track there Not bad for a model of the size All right, it was just so like low-drama I guess that I don't have much to say All right, well that is going to conclude our first look in test of 10 cent h y three

Now that it is out of preview, it is Apache 2.0. I think for the size of this model, this actually was fairly performant and will be very interesting to see how folks do with this. When it is quantized and able to run on 256 gig unified devices, whether that be a DJX Clark. DJX Spark cluster, a max studio, or some other mix up of dedicated VRAM and system RAM. So this could be very interesting for the hobbyist setup. I suppose for a quick results overview, it was interesting because the first C++ gate game was actually better than the second one as we can see here. There was an issue with the orientation of the player. That wasn't the case with the first result. Now I would say for the size of this model for a first result, it actually did quite respectively. So I was happy to see that. Additionally, what else did we have that was notable? Our skydiving game actually had some really decent looking... like... terrain to it. So if we look at it again and I'm not gonna go through the entire thing once more, but the terrain was really quite good. And this is one where I can definitely see a reasonable comparison to the GLM 5.2 result. It wasn't as good in terms of that one actually had some gameplay elements, so you had to fly through floating rings, but the terrain in this one was really much better than the GLM 5.2. So that was cool to see, at least if my recollection is correct, but unfortunately we had some visual oddities to when the parachute was actually deployed. Nonetheless though, it was pretty interesting. The city timeline result was definitely less good than the GLM5 2-1 was. I mean, it had some sound and it had some slight effects in the transposition between different eras. It just wasn't really that good. comparatively to the GLM-5, to result in some other ones I've seen. Overall though, I mean it was competent enough, everything worked. There were some different sounds between the era, there was changing of building heights and then things that were on the roofs of the buildings as well. And then it had a very optimistic and futuristic outlook of the year, 255. So hopefully this is kind of what we can look forward to in 30 or so years. Then we had our V8 engine test, which was just done so fast that it was quite impressive. The overall result was not perfect, but it was more or less in the style of a V8 engine, and it did include the hole for the 280 motor, which was interesting to see as well. And it did fix it rather quickly when we told that the motor orientation was incorrect. So it was just called to see in a different sort of test, which brings me into the other different tests that we did, which is fortunately not something that I can show once more, which was the... keyboard backlight fiasco where it just ended up actually adjusting the backlight for the system so it did show some level of being being able to work with like a lower level accessibility system hardware it just wasn't 100% on the ball with that And the entire cost of that test was 43 cents. So basically everything we saw done today for the entire duration of this video was 43 cents. So with that, that is going to conclude our first look and test of 10 cent HY3. I do want to just make sure one more time that this isn't on open router now because that would open up a lot of ease of use for folks. Of course it is. Of course it is there now and it's free as well. That's great. Well, nonetheless, you can now go play with this for free as well on OpenRouter which I would imagine will increase the amount of interest in this model. So with that, any questions, feel free to leave them in the comments and thanks for watching.