Tencent HY4 Is INSANE– Is THIS Tencent’s Next Frontier Model?

0:00 Let's check 1965. That is well done. Today we're going to be taking a look at a really cool new model from Tensson called HY4 Preview. Now, this is, as you may have guessed,

0:11 the preview version of the upcoming HY4 model. And over the course of the HY3 model from the preview to the full-fledged version, we saw some pretty significant capability increases. So,

0:22 having another almost frontier open-source player in the mix is quite exciting. Now, before we get into it, a couple of things. One, do feel free to subscribe so we can hit 75K, then 100K.

0:34 So, essentially, I've had a pretty difficult time actually just testing this in a smooth manner. So, keep that in mind. However, I'm really excited to play with this because there are some

0:42 interesting things about it. One of those, and something I do like is there's no like, hey, we're going to release the weights in a couple weeks. They've just slapped them here on

0:50 Hugging Face as well. And while this is a pretty beefy model coming in at 770 billion parameters, it is a mixture of experts. However, it has 49 billion active. So, this has a pretty heavy

1:01 active parameter count which hopefully will show some pretty good capability in terms of coding specific tasks. Also, it has a million token context window and they say it makes the biggest gains

1:11 where work is hardest, such as long horizon software engineering and document heavy office work. But I'm more interested in the software engineering stuff as it will showcase well on our

1:21 tests. So with that, let's take a look at some of the interesting things of note about HY4 and then we'll get into some fun and light testing. So let's start out just by taking a look at some

1:29 benchmark JPEGs. However, prior to that, I am actually going to initiate the browser OS test right now because it can take a little bit of time. I've currently been also fighting through

1:39 trying to get that whatever you say. Okay, thanks. That was Oh, because I gave it a little SAS first, but it gave some sass back. It was like I've not had one single piece of luck trying to use

1:50 this through open router. So, I'm just trying to get the browser OS from here. I don't think we're going to be able to. So, I've also reluctantly started it in Open Code. Hopefully, I don't hit my

1:58 pretty darn low limits with that. Let's take a look at these benchmark JPEGs. It's actually showing quite a leap in capability. And something I do specifically want to drive home is this

2:07 is currently in preview. Now, models that have been in preview previously were things like Deepseek V4 Flash. And when that came out of preview, it was an absolute monster. So, there are still

2:18 massive gains to be had from where this currently sits to when it's actually complete. So, the darker purple on top of the light blue is just how this is performing compared to the full-fledged

2:27 HY3. If this is really stacking up a 64.3 on Deep SWE, that does put it pretty comparable to some big heavy-hitting models as we can see right here. Additionally to that, this is just

2:38 showing that the capability leap between this and its predecessor across the board seems to be pretty significant. Let's just quickly take a look at some more pertinent information about this

2:47 model. And for this, we'll refer to the hugging face model card. Again, it's nice to see it is uploaded and posted here. The license is Apache 2.0, which is a very highly regarded permissive

2:57 license. So, it's nice to see it posted under that as well. And additionally, as we did see the size, a 770B mixture of experts with 49 active. This does also have a built-in MTP layer for

3:08 speculative decoding. So basically, it will just generate tokens faster, which is nice to see. The context length is about a million, which is a very healthy context length. And additionally to

3:18 that, they just mention again some of the things this is built for. And I do like seeing here game development listed prominently as one of the key features that this is built to be able to excel

3:27 at, just cuz that makes it exciting. Then again, because this is a preview, they reiterate that there are some known limitations. There's real headroom left in both pre-training and post-training.

3:37 So, they're shipping with known issues among them spending longer than necessary reasoning through complex tasks and a tend to overify tendency to oververify its own work. So, it may

3:47 become like necessary to stop it and just look at a result that's made and then give it the specific issue as opposed to having it try to reason over generations to try to find issues.

3:58 Sometimes that speed things up, especially with a preview model like this. Additionally, the FP8 model weights are open weighted as well. You can see them right here. And then they

4:07 just have a bit of information on how to run this locally for those that can. Now, in terms of pricing, it is not necessarily cheap, I think would be said. It's 83 cents per million in and

4:16 then $2.50 per million out. There is a lot of competition in this price range. Even things like the lowest tier GPT 5.6 model is significantly or at least a bit cheaper than this. And there's just a

4:28 bunch of models appearing now in this general pricing area, which is really nice and good to see. From what I just saw right there, this wrote 5,000 lines of code for the browser OS. Now, it's

4:39 currently looking here in the system just to see what autonomous things it can use to test this. I have half a mind to be like, okay, let's just pause you right here so we can look at it. It's

4:49 now definitely it's writing a script just to be able to test certain specific things. As we can see right there in the browser because of how restricted the limits are with Open Code Go. I'm almost

4:59 half tempted to just like test the script, pause this, and then if there are issues, describe it to them rather than do this. I'm actually going to do that because of the circumstances of

5:08 this model. The testing is not necessarily like fully reliable. So, I want to try to get more results as opposed to like single ones. So, let's just take a look at this. Keeping in

5:19 mind Nexus may have some issues but we'll fix them. Okay, overall so far this definitely looks pretty good. It's like a nextG result. I am going to ensure the speaker is turned on as well.

5:31 Synapse is a three layer kernel that every app is obliged to use. A shared virtual file system. Okay. Event bus set of background dammons. Nothing in Nexus OS is a self-contained widget. the mail

5:42 client, the terminal, the file manager, the wallpaper engine, the music sequencer, and both 3D games are all clients of the same system. Okay, not bad. And we do have a pretty interesting

5:51 background effect. Actually, like this is a fairly decently done background. Let's check the big thing. There is a right click. Yeah, good. And then we do have a bunch of things in the bottom

6:03 right such as the correct date and time in our local as well as uh Okay, so that's to shuffle wallpaper. We'll just pass on that. Good. Our start menu looks good. Okay, I typed H. I don't know like

6:15 how Neon Maze has H or Well, that's okay. The search bar does work. And I can't get my eyes off this synth one because that's always excites me when there's like musical things. In the

6:24 meantime, let's just start with our GTA clone. I think it spawned us in a building. We can see kind of by the mini map like what's going on. Unfortunately,

6:36 that's all right. Sometimes I don't think we weren't spawned in a building because when we move we could see this does seem to have some potential. So just chalk this to the list of issues

6:45 that we'll be able to manually tell it to fix. Next up, neon maze. This is actually not glitched out. I think it's just kind of dark and hard to see, especially

7:00 Oh, system failure. And then we got a notification there of new high score. Overall, not bad. And I am pretty eager to see the GTA game. So I will ensure we fix that and actually get a peek at it.

7:09 All right, we've got our mail app. Good. In this theme of the rest of this, yep, we do also all the models. It's interesting. They just try to like implement other things, meta references

7:20 to these apps. The maintenance window has moved. Drones are still aggressive in the maze sub levels. So I'd not go poking around till I've patched the targeting firmware. Check mail. Good. Oh

7:31 no. Okay. So, I was wondering if it would like um interact, you know what I mean? Like if we clicked check mail x number of times, would it then emulate a new mail? That is what happened. Hey,

7:44 this looks pretty good right here as a file system. And when I doubleclicked on the top bar right here, it actually expanded it as it would in a real life operating system. Nice attention to

7:52 detail. We also do, I believe, just saw some save game files or something for our games, which is cool. There's a good list of like emulated files here. All right, let's check our terminal. Kill

8:03 process ID. Okay, this is like a pretty This is a pretty intricate. Is there I just want to try Neoetch. Good. It's like a ghost or a Nexus portal. Okay. Of course, the Matrix.

8:21 Good. And it does it in the terminal instead. Let's see if it has an auto timeout just to show some attention to detail. Okay. I don't think it does. That's all

8:31 right. Synth. Oh, yeah. All right. Oh, look at this. All right. All right. All right. Good. So, that did actually take effect. The biggest issue is we don't actually

8:57 see it like move through each of these bars or whatever you want to call them. But based off of turning all the snares on and it properly reflecting that in the sound, it did in fact work.

9:11 And we have a sound bar here in the bottom right. Chip tune. Good. Just tell me now. No, I'm kidding. It's too early for that.

9:31 Save to disc settings. Adjustable accent colors. Okay. And they do reflect. Wallpaper. Nebula. Not bad. These wallpapers are actually quite well done.

9:45 Oh. All right. That one's ocean. Look at this. That's kind of cool. Ridgeline neon rain

10:01 hideous culture also kind of disturbing but like in a different way. We have our adjustable hue shift. I think I'm going to put this back to the this one because this was kind of cool. There's a

10:11 wallpapers thing. Oh, okay. We already kind of saw this, but it has its own dedicated app as well. Randomize. Would that actually change? No. Okay. Going to just ignore that. Then synapse our

10:24 special feature. Oh, it was just basically the fact that apps can interact with other portions of the OS. So when we had the synth playing when the background was actually moving to

10:32 reflect that is a demonstration of the special feature. So basically that interesting. All right. Really our only problem that was noticeable is the GTA game kind of didn't work. So I'm just

10:44 telling it to fix the GTA game and not to test it itself. I will test it manually and tell it if it works or not just in the interest of saving usage. All right. Supposedly, we have a fix for

10:54 the GTA game not working. So, let me just do a hard refresh and we'll see if it actually did. I'll make sure the sound is on as well. All right. Nice. Nice. That is well done. Look at that.

11:05 We've even got like this is very, very well done. Now, they did pretty prominently showcase this model with a lot of like aesthetically pleasing 3JS things in the announcement post. So,

11:20 okay. There are collision detectors on the vehicles. I don't see any pedestrians. Oh, yep. We do. This is really well done. I should full screen it. I consistently forget to do

11:36 that and I don't know what to make of that. And the sound of the sirens grows fainter as we get away. So, it actually has some like directional understanding to that.

11:47 All right. Well, that was not bad. Not bad at all. Actually, quite good. And it did fix the issue that we had that was causing it to not work. Since we often run it second in our testing, I think I

11:57 will just opt to do the self-contained C++ skateboard game here as well. This is the New York City map one, and it cannot use Ray Lib as this is a large enough model that it should be able to

12:07 do this without even if it is still in preview. We can also see in the meantime that it is still working on the browser OS here just from within um open code. Now I noticed that there were still some

12:18 upstream provider errors here but it should retry. So hopefully fingers crossed we'll be good to go. So the C++ gate game was going. It seems like the files written and it's just trying to

12:28 fix some issues. However, one of the biggest problems right here is what we can see where even when paying for this it's still just hitting rate limits pretty horrifically. So this is like

12:38 borderline unusable for any proper agentic tests like this because they require so much back and forth that it's hitting these errors pretty frequently kind of negating the work. So, the good

12:49 thing is we did have most of a file written, but it should hypothetically at least show us the first Oh, man. That is Skate NYC East Village 6:15 p.m. Oh, wow.

13:05 This looks potentially extremely extremely impressive except for the lighting issues right here. So, I'm actually controlling it. Look at the the building signs and things like that. So,

13:16 wow. Is that scaffolding for construction? I'm going to say this is like a result where we we can't even really see it and it looks super super impressive. So, what I'm going to do is

13:27 just stop this because this is just going to glitch out anyway. I'm going to tell it what's wrong and then hopefully we'll be able to have it fixed. All right. I've given it a pretty targeted

13:35 instruction to just fix the lighting and then say done master and to do nothing else because just trying to mitigate these rate limit errors. Okay, so it finished it and then said done, Master

13:47 Bjan. So instruction following we can say is pretty good. All right, let's take a look at this. That's really Is the sound on? Yeah, it is. East Village 6:15 p.m. There's smoke

13:59 coming out from the manhole covers. There's even water. Was that someone walking a dog? Is that a phone or was either a phone or a weapon? I'm going to assume it was a phone. Hold on.

14:10 I'm trying to look at this dog. This is like, oh, okay, we're stuck. No, we're not stuck. This looks really, again, we have to remember this is a preview, but if this is what the preview

14:23 result for this looks like, imagine what this is going to look like when it's actually completed. Oh, we have to kind of like P Oh, the sparks. Uh-oh. Okay. Yep. So, that is a

14:35 telltale sign of just like small model smell when they snap to the rail like that and unfortunately get stuck. However, I will be a little more forgiving due to the fact that this is,

14:45 of course, a preview. The cars are not all there in terms of their drawing. But let's see if we can get straight onto that rail. Actually, I don't think we want to.

14:56 This is cool. There's even a slight like pump effect, which is when you're skating and you push the board with one of your feet so you can actually move. I swear we just got hit by that vehicle.

15:06 Let's see if that still to this date, I believe, the Quen 3.8 8 max model did the best on this. However, that was a north of two trillion parameter model. This is 770

15:19 and still in preview. There's a hot dog vendor. This is a very very very very promising result. I will say and I'm pretty darn excited now to see what this ends up becoming when it's out of

15:30 preview. Just having more models that are actually like quite competent is very exciting. We have store names as well. It seems like it didn't properly draw the text that would show up like

15:40 for the scores and things like that, but I'll forgive it. This is very, very, very well done. Well, for a preview, it's definitely previewesque, but it's well done. So, in the meantime, I had

15:51 also given this the Subway FPS test, but I did it through Open Router just because I wanted to see if we could get some like parallelization to this test. And I said, just give me the single file

16:02 as the deliverable. Do not try to verify it. Just give me the script. or at least that's what I assumed. Yeah. Okay. So, that's not expanded. And it seems to have actually properly adhered to our

16:11 instructions. So, we have a script right here. It does not look very long, but I'm interested in just seeing whatever we get because just from inline, keep in mind this didn't do this agentically. It

16:22 just generated the script based off of its thought process. So, we'll see what we get. Subway panic. Okay. It is possible that the open router result is uh okay. I think maybe we'll just

16:36 let's not include that. I'm now wondering if there was some like uh thinking setting that was not properly set here. However, it does show that we're able to get at least some simple

16:47 results just from this in line through open router. Uh I think maybe we'll try this again more agentically cuz I'm curious. Yeah. So, I'm giving it the subway test, but just from within pi

16:58 right here. Oh, okay. We'll see if we're even able to get anything for this. If that's happening straight out of the gate, I'm not feeling 100% bullish that this will work. But we'll just kind of

17:07 let it limp along in the background and see what we get. All right, it's at the point now where it's trying to smoke test, which basically means that for all intents and purposes, it's done writing

17:16 the script. Now, I may be a little eager to just look at this because to be honest with you, look at the amount of failures that have been occurring here. And it's really quite frustrating

17:25 because to even use the free tier they're offering on Tencent Cloud here. I had to put in my credit card info and authenticate with a phone number and things like that.

17:34 So I'm a little frustrated. This is just garbage serving experience. I will say the model being served here is garbage. Now the flip side of that, it's Apache 2.0 open weight model. So it is possible

17:45 to just like pop it on RunPod and not have to deal with this stuff. But with that, let's take a look and see what the Subway FPS result has given us. All right, we have our Meridian subway game.

17:55 Good. Good. Brightness, sensitivity, volume, quality. Interesting. I don't know that I've ever seen like a graphic quality setting. Holy. All right. Oh, it's frozen. Sometimes it

18:08 takes a second to become unfrozen. Hopefully. All right. This is really, really good. I mean, some of the

18:21 textures here are reminiscent of the Opus 5 freak result for this. Hey, look at that. We actually have a digital station like sign here. Now, one of the big issues I'm noticing,

18:33 well, one of two. Um, the like crosshairs are not necessarily where the weapon is pointing. Also, I don't see any enemies. For preview, this is quite good.

18:49 We have a spill. Even like a cup there. Like the big problem is there are no enemies. Control is to crouch. We can also swap weapons. Right mouse button to aim.

19:04 Oh, look at that. All right. Two. Oh, wow. And then control was to crouch. This is cool. Mind the gap. Oh, the big issue is that and then there

19:22 are no enemies. But as a preview result, this has a large large large amount of potential. Platform 2. I'm very excited for full-fledged HY4. Next up, I'm going to be giving this the

19:36 beautiful watch website test where it needs to have the 3D model of the watch, the cinematic camera effect, and a keyshot style render. However, I've also added something in just to make this a

19:46 little more immersive. It needs to include an exploded view of the watch controlled by a slider in a dedicated section of the page. So, ideally, there'd be a slider and when we moved

19:55 it, the exploded view will expand or contract based on where the slider is just to kind of keep something that will give us reference point to other models, but add in some things to make it also

20:05 not so like the same. It's not fully done yet, but after 307,000 tokens through open code, I am curious just to see what we've received for our watch effect. Darn it. Okay, so the big issue

20:19 here is whatever this black square is that's moving around kind of preventing us from really seeing this watch. From what we can see in the corner here, it does look quite quite quite. Yes, there

20:30 we go. That is Look at that. There's actually Hey, get out of here. There's Okay, so Oh no, there is a date wheel and it actually shows the dates that are upcoming and also like there. Okay, so a

20:47 bit troubled, but it's been that's cool right there. It's been going on for so long and with the 429 errors, this is basically like unusable for anything real. So the fact that we've received

20:58 this is good in and of itself. All right, I fixed this manually just so we could properly see it. At least I think I did. Okay, I think I did. This looks really darn good, except for the weird

21:08 seam there. It's actually drawn some like leather style material on the band, which you can see at certain angles. Additionally to that, this was a cool piece right here just showing inside of

21:19 it with the movement actually spinning and working. Something cool to see. Then we have the renders of the two different style ones. Both not bad. Now, the final thing that we were not really able to

21:30 see before is the exploded view. So, as we drag this, it should do an exploded view of it. And it absolutely Oh, look at that. It actually put in like little tool tips of each one of these things. I

21:41 didn't tell it to do that. So, that is really a nice nice touch. I'd say that right there. That is quite well done. I'm quite pleased with that. Especially the tool

21:52 tips that it added in, which it totally didn't even need to. So, it went above and beyond for that. 12 components, one heartbeat. Apparently, we can also change the specific style of the watch

22:02 and there is even some movement emulated right there and we can change it to the other style watch as well. So, overall I would say that right there, somebody in the comments I believe had suggested

22:13 this as an additional piece to this test. Very, very, very good. And it's got some like smooth movement too. Even if I just like slam this left or right, it still is like buffered in terms of

22:23 resting place. Very well done. Fortunately, I had decided last night to do a couple of tests, anticipating that there may be some problems with 429 errors and rate limiting. One of those

22:34 was the Chrono City test where it needs to showcase a 3D city block over a number of different eras and we can use a timeline slider to select which era we are going to visit. Now, interestingly,

22:45 this opted to do this using Blender. The prompt does not specifically denote any language that it should use or any method. So, it found that Blender was available and opted to do it that way,

22:55 which will make for a pretty interesting view of just how this is able to work with Blender. It used the MCP for it. And it just gives us the script right here to actually open the scene. So,

23:04 let's take a peek at it. All right. So, here is our Krono City scene. Now, I don't 100% actually know how to work with this play animation. Okay.

23:20 I don't think there's any sound and I'm going to have some difficulty just navigating through this. Okay, so this is showing us city era 1945. It's interesting because I don't

23:32 specifically know how we would use this as a result like showing the timeline and being able to jump around. I honestly think I Oh, okay. The pedestrians are a bit messed up. So, I

23:44 do believe I need to do a bit of work here just to ensure that the right file and everything, but actually pretty darn cool. I mean, look at the buildings. We can even kind of see inside of them the

23:54 way they're drawn, the vehicle models and everything. This is a pretty dense Oh, okay. Now, we see the full Blender scene has loaded in along with sound. Now, this is 1945,

24:11 so I would imagine that it's uh I need to just be able to move around. This looks pretty impressive. There's a street car. There's sound design.

24:31 There's car models. Now, these don't necessarily look very 1940s, but we have pedestrians, though they are upside down. This is almost like overbaked in terms of the materials. I think though,

24:43 look at the emissive lighting from the storefronts and things. It looks pretty impressive. Let's see if can we figure out how to Let's check 1965.

24:54 That is well done. that transition effect, the sound, the vehicles have brighter colors and more in line with I think what we would expect from the '60s. Overall though, the scene is still

25:04 a bit grainy. I wonder if that's something that will become less so as it gets to the present moment, but still like kind of interesting. Let's All right, let's do it from this angle and

25:16 just change 85. Okay, so 1985. Unfortunately, there are no vehicles. Uh, let's go 2005. Still, this is I think this is pretty darn impressive the

25:29 way it went about doing this. Okay, 2005. I still unfortunately do not see any vehicles. Let's try 2025. Oh, cool. Look at that. The buildings

25:45 actually got way higher. Even though unfortunately, as we're seeing right here, things are perhaps not 100% together. Let's look at 2055. The buildings got even higher,

25:58 but unfortunately there seems to be some misisconfiguration or something. If we go back to 1965, that scene is more together. So interesting. Definitely previewesque, I

26:12 could say. Um, the NPCs are upside down, which makes for an interesting. Still though, I would say overall, I mean, even though this is a bit troubled for a preview model and for having used

26:27 Blender for the entire thing, that's pretty darn cool. And we've never seen this done through Blender before. So, overall, very, very, I think, good. But, let me know down

26:40 below. know, let me know if this is good or not, cuz I think it is, but it's definitely got a few issues. So, the final thing we're going to take a look at, and something that I did run last

26:50 night just in the anticipation of results taking a while, is the retro wrestling game. Now, I do believe this went ahead and just made this all in an HTML file. So, we've used 3JS for the

27:01 assets. We can see right here, this used 355,000 total tokens for this task, which ran for quite a bit of time. And this was just using it through open code. So

27:11 basically it says you don't need a server or anything. You can just click it and start it. So let's see what the retro wrestling game for this looks like. All right. I'm supposed to open

27:19 this one, not index.html. So that makes more sense. All right. Music on TWWA Titan Wrestling Alliance. All right. We have our controls and I just basically put a mash YUIP. Awesome.

27:38 Good. I like that because that's probably what I'm going to do anyway. Let's enter the ring. Good. Good. Jack Sunset Malone, Big Tex, Bull Calhoun, Nikolai Steel, and Raven Vega. Um, sure.

27:51 Ring the bell. All right. This This is really the way that it actually drew these body models, considering it wasn't allowed to use Blender or anything like that, is I

28:02 would say actually really really impressive. Detroit demolition. Okay, we got thrown out of the ring. These are always really really difficult to play. I've not had one that's like

28:18 I think we're stuck. I don't know how to get back in the ring. Look, it even put the the the Texan guy in a cowboy hat. Oh, okay. What are you doing? There is no sound. Unfortunately, the

28:33 sound effects are not working. Oh, look. Look. Oh. if these animations were smoother. I don't know what's going on, but I do have to believe that we are absolutely putting the Mac down on these

28:43 additional opponents right here. Okay. Yeah, we are. Cuz look at the health bar. Press P for the finisher. Texas tornado slam. It basically said just mash these like

28:59 keys. So, that's exactly what I'm doing. Okay, that guy has a chair, which could be a little problematic. Okay. Yep. Oh, this is quite well done considering that it was not allowed to use Blender for

29:14 these assets. So, the way it actually really using 3JS made these people have muscles and like hats and things relating to their specific theme of them being a wrestler. This is well well

29:25 done. And I like that we can just basically mash Oh, what the heck? Why did they just get more health? Press P. Finisher ready. Texas tornado slam. Okay, we get a closeup. I see like stars

29:38 over the opponent's head there. He just got laid out. Raven Vega is out. All right, cool. Oh, what the heck? They put someone else in with full health.

29:55 All right, well, this was good. And I think the fact that I've probably spent the longest time playing this out of all of the other wrestling results we've received probably speaks volumes about

30:04 how this is a good result. So overall, that's probably going to be the final test that we run with this just due to the fact that and for some closing thoughts. Unfortunately, this was one of

30:16 the worst experiences I've had and actually just trying to test a model. the 429 upstream rate limit errors were just really really bad to the point where I gave up and gave them my credit

30:26 card number and phone number and things like that to be able to use this and it was still giving me those errors arguably worse than I was getting them just through open code go. So that is

30:35 somewhat infuriating. However, overall for a preview model this thing seems like when it is out of preview and actually fully baked, it is going to be an absolute monster specifically in like

30:45 3D game generation and things like that. So, I'm going to be extremely extremely excited to test this one. It's out of preview. I think this is something that is going to come out hitting quite quite

30:55 heavy comparatively to other models that are available at that time. And it's Apache 2.0. It's not horrifically large. So, this is a similar overall footprint to GLM 5.3, the full model. And if they

31:09 came out with a smaller version of HY4, I think that would be like a flash model or something, that would be absolutely fantastic. So, that's something I want to see as well. I suppose we can do a

31:19 quick results overview as well, which I guess I'll stop playing this game and do that now. But, this was really, really, really enjoyable to play. All right, so we had that result, which was obviously

31:27 very good. The watch website was awesome. And not because the watches were like the best material-wise and things like that, but the idea right here to actually include the exploded

31:38 view and then the fact that it went ahead and put tool tips and references to all the individual pieces down here. This was very nicely done and very smooth. So, I was really happy with

31:46 this. And the suggestion to do an exploded view as well, just to kind of freshen up this test a little is something I'm very inclined to keep around. It's just kind of neat. Our

31:55 skate game was again one that for a preview model, like look at how the newspaper was actually floating there. Is very promising to see how this evolves when we get the full model. We

32:05 had bail mechanics. We even had sparks when we were grinding. There was also an NPC walking a dog and then what looked like Oh, look. There's two dogs being walked right there. Oh, and we slammed

32:14 into the hot dog cart. But they also one of them had a phone in their hand and this was just actually I don't know if that would be a phone cuz this is an early 2000s map, but still that's

32:24 probably beside the point. And we could pump to Oi. The tricks worked well and the board did move separately from the player model. So that's always something to check. And this was just a proper

32:35 proper result for a preview. So, our subway scene as well was pretty promising, even though unfortunately it was a bit glitchy when we first got a look at it. And then ultimately, there

32:45 were actually no enemies here that we could interact with. That one was very good. And that brings me also to the next thing right here, which was our Blender scene. Now, it was really

32:54 interesting that it did the Chrono City from within Blender. This was again very promising and just seeing how it was able to actually interact with an MCP. The transitions to different eras were

33:04 good. There was sound here as well. That's so so so frustrating now that for some reason this is showing us vehicles in the 1985 scene. It was not doing this when we actually tested it. Okay, 2005

33:18 is also visible. The buildings get taller and then 255 finally. So, it's odd how for some reason some of the vehicles were spawn. Look at that. That scene right

33:30 there is cool. I know it's super grainy and things like that, but this definitely showed some promise. And I'm pretty excited to do some more Blender and like GDAU testing with this thing

33:40 when it's fully out of preview cuz the potential here is vast. Look at the the hover train. That's just like that's awesome. And back to the subway station scene that we were taking a look

33:53 at. This was also very good. We had multiple different weapons. We actually had the ability to aim, the ability to crouch as well. Unfortunately, two things. bullet holes did not go in the

34:01 environment and then the enemies didn't load as well. However, the textures here looked pretty darn good. Even just putting in like a random tile like that that's more broken up. It's just

34:10 attention to detail that's very nice to see, I would say. Then finally, we of course had our browser OS. Overall, it was good. We needed a fix for the GTA game to work, but once that was

34:19 implemented, it was overall pretty competently put together result. A big city, nice map, buildings were drawn well. It was not super exciting when we basically ran into someone. However,

34:31 more or less a good result, especially for a preview. And the wallpapers were funky and the synth as well was a nice touch. So, overall, that is probably going to conclude our first look and

34:41 test of the Tencent HY4 model, which we can see right here is still in preview and will hopefully sooner than later be out of preview. And again, I'm really, really, really looking forward to

34:52 testing this once its full-fledged release has appeared. I also would be very very very excited to see if they would come out with a flash model or something a bit smaller that could

35:02 realistically maybe fit on one or two DGX Sparks kind of like GLM53 flash. So that I think would be a really potent competitor and something a lot of folks would be eager to see as well. So let us

35:13 know in the comments if you No, I'm kidding. All right, so that's going to conclude our first look and test of Tensson HY4. If you have any questions, please feel free to leave them in the

35:22 comments and thanks for watching.