Ornith 1.5 35B is INSANE - Full Test & Quant Comparison!

0:00 It made these assets. It put it all together. Definitely gelatinous. Today, we're going to be taking a look at the newly releasleased Ornith 1.5 family of models. In specific for this

0:13 video, we're going to be testing both the Q4 and Q8 quantizations of Ornith 1.535B- A3B. Now, being that there is not currently a Quen 3.835B- 835B- A3B. It

0:27 is possible that this could be a suitable replacement for the time being, assuming that does eventually come out, but Ornith and their previous 1.0 models, I did cover some of them on the

0:37 channel, and they were truly pretty potent in terms of a capability increase. Now, for a bit of background information about these models, they are essentially, as we see right here,

0:46 developed on top of Quen 3.5 for this one in specific. And they are trained additionally, and it seems like the training that they do for them has some pretty potent capability increases. And

0:56 a lot of folks are definitely interested in more models that can run on consumer hardware that is actually reasonable to get your hands on. So, before we get into it, do feel free to subscribe so we

1:05 can get to 75K, which is a good milestone to get to 100K, to get the plaque. And with that, let's start out by taking a look at a few interesting things about these models. So, for the

1:15 release today, while we are going to be focused solely on the 35B mixture of experts model, they did also release a 1.5 397B, which is apparently here scoring a 56 on Deepsw S., which if that

1:28 is true, that's actually pretty darn impressive considering, as they say right here, that matches Claude Opus 4.8. Now again, take all of these with a grain of salt just because you never

1:38 know with benchmark JPEGs. It's always best to like run your own or see something of the sort, but it does seem like overall these are all a pretty big leap in capability over the previous

1:46 versions. And then the model that they are beating out, which is just the base Quen 3.6 variant of the model for today's test in specific 35B- A3B. So really without getting into too much

1:58 else, I do just want to mention the testing setup real quick. So, I'm going to be opting to use these both side by side from within the Pi harness. On the left here, we have the 35B- A3B at a

2:09 Q4KM quantization. This is just the GGUFS that Ornith themselves have uploaded on Hugging Face. So, I'm using their own GGUFS. And on the right here, we're using PI as well. And this is the

2:21 Q8 quantization again from Ornith themselves in terms of the GGUF. Now in terms of testing setup and methodology and everything of that sort, everything that is listed here in terms of

2:31 suggested sampling parameters, all the reasoning like special things that need to be properly implemented here are. So no worries about like wrong sampling parameters or anything on the back end.

2:42 So in terms of the system that these are both being run on, I'm actually running both of them on a 6000 Pro card at the same time. So, we're going to get started with the browser OS test v2.7,

2:52 which is the same as the previous one, except they now need to include procedurally generated wallpapers with movement. Additionally to that, they need to also have an email app, which is

3:01 like a micro creative writing test cuz sometimes they include like interesting little emails in there. First to cross the finish line. No, our Q4 finished first. As we can see right

3:13 here in the left pane, they were both just autonomously checking over. We can see here just throughout the duration of what it was doing, we have a decent bit of thinking, but because these models

3:22 are so quick on this specific setup, it doesn't really present itself as much of like a patience issue, at least to me. So, additionally to that, there were some errors. I noticed that it would try

3:31 to make some changes and be like, "Oh, I need to just do this like in small pieces instead because it was not able to make large scale changes like at one time." So, it broke that down and then

3:40 checked it over a few times, had some issues, fixed them, and now we have our result while the Q8 is still working. All right, so we can see here the Q8 is hanging. I think essentially to put it

3:50 simply, the tests that it wrote passed, but it mistakenly like ended this with close the tab when it should have ended it with close the browser. So basically, like the thing that would kick off this

4:02 to be like, hey, it worked like everything's cool is never going to come through. So I have to manually just stop that and be like continue. Yeah, we'll see if it catches up. The test aborted.

4:12 Let me run again. Capturing full output. Okay, that's going to do the same thing. This is maybe a bit of help, but that's okay. It, you know, that's the issue that it was having right there. So now

4:24 it should properly handle this unlike before. So after quite a bit more time, we now have both results. The Q8 has finished up. it did a lot more intricate like QA

4:36 checking essentially. And we can see here the context lengths are set the same for either of these specific models. So the Q8 used up 41% of its 262K context window and the Q4 used up

4:48 only 19 12%. So with that, let's now take a peek at the results. They are in their own folders or else it would get perhaps a bit confusing. I think I'd like to begin just by taking a look at

4:58 the Q4. All right, so here's our Q4 result. They both named it like pretty much the same thing, which I guess is to be expected. Okay, overall not bad. And I do recall in the summary of features

5:09 it added. It mentioned that there was actually a voice assistant, so that's something to try as well. Okay, there is no right click, but we do have the time and date in the bottom right. I will

5:19 hold off on trying the voice assistant as that is the special feature. Our start menu. Okay, good. That does work. Lock screen unlock. Okay, nice little feature. Nonetheless, let's just start

5:30 out with our GTA clone. Okay. And we have a pretty clear error right here. So, at least that's good to see. Can we minimize this and reopen it? Okay, we can. I would imagine that our void

5:41 runner game will work fine is what I was going to say. Okay, but we can't full screen it. It just stays a fixed size, but that's okay. Can we shoot or do anything space to fire? It's a weird

5:53 noise. I don't see like firing. So, but it actually did work. One of the games did work, so that's nice to see. Let's check our mail app. Okay, this looks very Mac

6:08 OS style. Welcome. Delete. Cool. City services. Our cameras noticed some enthusiastic driving in Mega City. Please drive safely. That's concerning. All right. Compose.

6:19 Solid. Next up, we have notes. Uh, okay. It autosaves. It's, you know, functional. Maybe not the best looking thing, but good. And the full screen worked properly. Settings. So, this is

6:31 where we'll we'll see if our wallpaper is going to work or not. Okay. Uh, we can forget what I was just about to say. Then, finally, terminal does seem functional. Oh, I clicked out of it. All

6:42 right. Let's try our special feature now. Oh, okay. All right. So, to try our special feature, I just had to swap to Chrome. Uh, I was using something else.

6:52 That's all right. Hello, >> I am Arya, your Nova OS voice assistant. I can open apps, change wallpapers, lock the screen, and more.

7:03 >> I am Arya, your Nova OS voice assistant. I think you may have heard apps, change wallpapers. I gathered that. Thank you. >> I am Arya, your Nova OS. >> Okay, I'm I'm well aware now.

7:15 >> Still, that was actually open void runner. >> Sorry, I didn't understand. Open VoidRunner. Type help in the terminal for commands.

7:25 >> Sorry, I didn't >> open mailbox >> object html div element. >> Good. Good. All right. Next up, I guess we'll just look at them in Chrome for

7:34 now. Here is the Q8 result, I would imagine, because this did a lot more intricate troubleshoot. Okay. Yep. We can definitely see a pretty significant difference between the two of these.

7:44 They both did put in this weird little assistant. I'm going to close that for now. This is a Okay, there's no right click, but there is significantly more here in the bottom right. This is a more

7:52 stylized result. And this is pretty blatantly just a difference between the two. Let's check our start. We do have that screen lock thing as well because we do see some similarities just in the

8:03 names they choose and things of the sort. Let's start with our GTA clone that is quite quite um pulsing. It's quite good, I think.

8:17 It's actually like very large. Shift to sprint and boost. I wasn't pressing that. What happens if we do? Oh, wow. Okay, this could be like a marathon simulator. That's not a bad idea. Can I

8:30 borrow vehicle if I can get near one? Get back here. Why even bother? I mean, we're faster than them. I don't think that works either. There are no mesh colliders on the buildings, but I'm just

8:40 going to try to get to this car. What about the orange one? Okay, that's fine. I don't like that color anyway. All right. Well, it's definitely gelatinous.

8:54 Definitely better. All right. Next up, Starborn, which did work with the other result. Yeah, equally as, you know, uninspired of an idea for a second game that we're used to. Also, in some ways,

9:08 this one is almost not as good because we don't actually see a ship male. Okay. Bank statement archive. Does it actually go into archive? Okay. I I'm I dig that. That

9:22 was pretty cool. Marcus lunch. Oh, so this space bar doesn't work in here. All right. Well, Marcus will understand. Good. Next up, Aurora. Is this like a wallpaper thing? Oh, no.

9:36 This is the special feature. We'll skip that for now. Does this save it to a txt? No, it saves it to browser storage, which is acceptable. Settings, aquade,

9:46 deep, sunset, nebula, ocean. Oh, good. Wallpaper motion speed. Did that make a difference? Yeah, it did. Taskbar opacity. That's actually cool. Look at that. Clock format 24hour.

10:00 It does. This thing's pretty good. Now, this test has been around for a while. I need to change the wallpaper frequency or else it's going to be highly, highly distracting.

10:11 A well implemented matrix feature. Can I get out of it though? A smart model would time out after a few seconds. Good. I was happy to see that. Photos. Oh, wow.

10:22 All right. Then finally, we have Aurora. Tell me a joke. Oh, that's some Okay. If this says the space bar doesn't work, so it's I told my CPU a joke about bandwidth. It

10:35 didn't get the bandwidth. That one's actually kind of clever. All right, this was definitely a significantly better result. I'm just going to be completely honest. So Q8

10:48 here for this side by side was definitely definitely the way to go if possible. And being ane, I mean, it's more possible to kind of offload a lot of this to CPU RAM if you have the

10:59 available CPU RAM and it'll be slower, but the quality may be pretty significantly improved just based off of this one sample size. So, with that, let's do something else. Next up, we're

11:10 going to be giving both of these the Subway Simulator/FPS prompt. This will be one that I like doing side by side because it'll again be like the browser OS where we'll

11:20 likely notice a pretty stark difference in capability between the two and it's something that can be run side by side. Following this, I would like to attempt C++. And for that, we'll likely just opt

11:30 to run one at a time because they'll want to like run it and troubleshoot and I don't want to confuse or overload them. Both have now completed and again the

11:39 context length used up by the Q8 was significantly higher than that of the Q4. It basically spent a lot of additional time just ensuring that everything was 100% functional at least

11:49 in terms of its assessment of the code. So with that, let's go ahead now and take a peek at both of these. Wow. Okay, we see some screenshots here maybe spoiling some of what we're going to

11:57 see. Sorry, I just got a little angry. Nonetheless, let's take a look right here at our I believe it gave me specific instructions that I'm definitely not following right now. So,

12:07 let's do that instead. Let's take a look at our That's great. Okay, so it's already running. Here's our Q4 version. Okay, I will turn my speaker on. We have explore mode and combat mode.

12:24 I'm just I hear things, but I don't see things, which is a little frustrating. Okay, let's just try explore station. All right, and again, this is the Q4 result.

12:41 So, okay, I'm going to mute this. You know what I think I'm going to do just out of pure curiosity, if nothing more, is just give this the developer tool console right here. All right, pretty

12:54 quickly. It seems like we have a fixed version hypothetically of the Q4. Okay, that's what we have right here. Interesting. This is actually passable. Now, the big problem is you're going to

13:09 be like, dude, like spin so you can actually like see. It's impossible. So, hold on. I can't actually move the view. It's basically just like a static pointed

13:21 forward. So, we'll try to assess this. It's interesting in some ways. This is no. Okay. So, it just seems like the mouse won't actually log any key presses

13:35 or clicks, I should say. The big issue with the Q4 was just that we couldn't actually move the camera around with the mouse and shooting didn't work. Okay, that's a bit

13:50 upsetting. I'm not noticing any fix. Let me just Let's try the explore station mode. No, it's still the same thing. Unfortunately, uh the mouse just doesn't

14:02 seem to affect anything. So, uh yeah. Try one more with enter combat. No. Okay. Well, interesting. All right. So for the Q8, I just followed up with,

14:17 "Bro, how do I run this?" Because I kind of left off not on really like a satisfying summary of like, "Hey, here's everything." So with that, now let's take a peek at our Q8 result. We can see

14:26 here just based off of the actual file structure itself. It definitely seems to be a bit more intricate. All right, so here is our Q8 meatball station.

14:40 I don't quite know how it came to that name, but I will say that Meatball Station's pretty decent for a model of this size. Okay, now unfortunately something that's gone wrong is when I

14:53 press the W key, the entire scene basically freezes. And when I let go of it, we haven't actually moved anywhere. Oh, wow. Okay, there are definitely some timing problems here,

15:06 but functionality is basic, albeit there, and unfortunately, no movement is capable. But if we are just to judge the scene as it is right here, are there bullet holes in the environment? I can't

15:19 actually tell because I don't think Oh, okay. It's very like what was that game where was it called Redhot or something? I think it was for VR where like things

15:33 kind of stayed in the environment like that. This definitely reminds me of that just by what we can see right here. Okay, so this is good. These are functional and we can actually exit out

15:43 of pointer lock to get to those and then get back in. So really the biggest issue is that none of the movement works and there's some like timing or like just some oddities to

15:54 like how the ammunition tracers just stay and don't go away and then as those blood splatters are very ominous. All right, both have now hypothetically fixed some of the issues we were

16:05 noticing. So first we'll try meatball station or the Q8 result where basically movement broke the game. Very good. And also one of the issues that we were noticing were that basically like

16:15 ammunition tracers would just stay in the environment frozen. Okay, those do fade away now, but we still have some of the odd behavior of the uh muzzle flash just never going away, but definitely an

16:30 improvement. Oh, reload. Good. And being that I never actually mentioned that as a specific issue I wanted remedied, I would say that this

16:42 properly fixed everything that we had pointed out. There are no bullet holes in the environment though, so that would be a complete failure that No, I'm kidding. All right, let's just quickly

16:50 explore this station while dodging these individuals. There are quite a lot of them. And we do have jump. Okay, cool. All right, so that was our Q8. And it did

17:03 properly kind of fix it and make it better. So, next up, we're going to run these one at a time. This is the self-contained C++ gate game test. This is the one where they have to make it a

17:12 New York City block as the map. And I have not included that they can't use ray. Whatever it is that they decide to use in order to try to build this, I'm totally cool with being that the I mean,

17:22 it's a 35b with 3 billion active at a Q4 quant. So, if it spits anything out functional, I will be quite happy. And I'm running these one at a time just because uh I don't want things to get

17:34 confusing if they're both trying to autonomously launch the games and test them. There are a little more like things going on here. So, this has been running for quite a while in the skate

17:43 game and this is only the Q4KM model that's running on the GPU. Nothing else is happening. So, basically, it's been putting out a ton of tokens. We can also see the context length is pretty much

17:53 close to just getting completely full. Now, at this point, I think it's probably reasonable to stop this as if it's not capable of doing this in this amount of time, I don't believe that it

18:04 is going to finish. It's been long enough, basically. Now, here's what I do want to do. I'm going to say just give me the file now. We'll see what it has. I don't actually think that's going to

18:14 even really do anything. And just for reference, here is what the Q4 had produced. I just ate my words so bad. This is significantly better than I

18:31 would have expected from this model. Now, yeah, it's kind of like Bunk, but like we actually have like a halfdecent rider model. Okay, I maintain that like maybe like

18:47 seeing this through all the way was a bit outside the scope of the video. However, the fact remains the Q4KM did this and like I just didn't expect it to even have anything that was that

18:57 competently put together. Sure, there were plenty of issues, but the fact remains there were elements there that were deeply impressive considering what one would have maybe expected from a

19:07 Q4KM mixture of experts model. Let's try this now with the Q8 and we'll see what we notice that may or may not be different. We now have our Q8's gate game result. This worked for I don't

19:17 even know how long cuz I went to bed. Okay. H, that's actually kind of cool. Now, yes. Uh, okay. So, there's no movement, unfortunately. I just want to make sure the keyboard's plugged in.

19:30 Yeah, it is. Can we Okay, so here's the thing. This is actually a pretty darn decent result for a model of this size, especially. It worked extremely extremely long. I

19:45 mean, the amount of like context compactions that occurred here, there were a lot of them. So, the fact that it even made something somewhat like as a basis, I think is pretty darn

19:56 impressive. Actually, I am quite quite pleased with what I'm seeing here. Even if there's no like real functional gameplay, there are NPCs walking around. There's a pretty funky and dare I say

20:05 cool like fountain effect. And the tricks do actually work to some degree where we can just see like the board. I'm just trying some of the traditional like trick keys to see if we can get a

20:16 kick flip or something. Oh, C is for different camera views and that works. This is pretty it's pretty impressive all things considered. I don't know what just happened. Oh, R

20:29 is reset. Okay, that's what I was clicking there. All right, so the camera button works. Unfortunately, no movement, but we can also reset. So, we do have some playability. Well, at least

20:39 in terms of like an OE simulator. I'm impressed. So, for the next test, and I did already run this one with the Q4KM. Interesting. Please connect Blender so I can start modeling. Okay, while Blender

20:50 is on this system, I do not want it to use that. I wanted to do what the Q4 did where it just basically used OpenCAD. I'm not even going to say it's installed on the system. It will figure that out

21:00 on its own. And I have not yet looked at what the Q4KM did for this prompt. So, this is, of course, I should probably mention the inline six-cylinder engine that needs to be modeled in a way that

21:09 is realistic to 3D print. The engine model is done. No, I'm sorry. I got to keep myself entertained cuz I basically just play guitar throughout the entirety of like

21:20 when these things are working. So, the plus side is I'm getting better at guitar. The downside is my fingers are sore. All right, so with that, we now have both of our engine models. And

21:30 let's begin with the small Q4KM. Okay, we do have a render which does show what may be reminiscent of an engine block. I think I'm going to opt to just look at the entirety of the

21:43 assembly as a whole first in a slicer program. It is engine blockish, I guess could be said. Definitely, we're going to notice. Yeah, not quite. Now, there is 1 2 3 4 5 6. So, it understood that

21:57 does like have some reference to, okay, it's a six-cylinder engine. There's 1 2 3 4 five of these. Let's look and see some of the individual pieces cuz sometimes we'll see detail that we

22:08 wouldn't have expected. Okay, so it's pretty much a very similar story that we're seeing right here when we open the individual valve cover. Let's just try opening the engine. So essentially kind

22:20 of the same thing, but it does show that it got the overall shape and some like info like, oh, there's six individual cylinder. So with that, let's now just look at the Q8 version. We'll just take

22:30 a look at the fully assembled sliced version first and foremost. Oh wow, this I think went ahead and just made like a full-size realistic engine because that build plate right there is of like a

22:41 normal size 3D printer. So this would extend far beyond it. Yeah. Let's look at these parts individually, maybe. Is there any text or anything? No. Okay. And this turns

22:52 into like this weird striped pattern that happens when it means it's too large to fit on the specific build plate for the printer. That's an end cap. Okay. There's two of them. So, we see in

23:03 some ways it the argument could be made that this is actually a worse result because the scaling is just so wrong compared to what the Q4KM did where it wasn't great, but at least the scale was

23:13 definitely more in line with what we wanted. I think this may be an attempt at like the Yeah. Okay. So, that's the intake. We'll have the oil pan. Okay.

23:27 Pulley, turbochargers, and it did understand in plural and this engine does have two turbochargers. So, just seeing that as well. It's a bit of pertinent

23:39 information. It may have web search to find that out though. So, I guess we can't definitively state that may be like inbuilt knowledge. Okay. And then right here we have the valve cover,

23:47 which is actually not bad. Again, the scale is just ridiculously oversized. However, 1 2 3 4 5 6 and those would be like ignition coil caps or something. So,

23:59 it's interesting to see the differences. This definitely had more depth to the level of detail that it opted to include in this model. The downside of that being for some reason the scaling was

24:10 just really really messed up and it even did understand here like that the N20 is only 20 mm x 41. So 25.4 mm is an inch. So we can see the scale there was definitely a bit off. Next up we're

24:23 going to be giving these the beautiful watch website front-end test which also must include a good-looking 3D model of a watch with a cinematic camera orbit over it in the hero section of the

24:33 website. So, this will just give us a good head-to-head of how their web design and some of their like model capabilities in 3JS are. Both of our sites have completed. Let's start by

24:41 taking a look at the Q4KM result. Okay, actually acceptable. It's definitely got some issues, but I will say some of the core components are there like the crown. The watch dial is

24:56 somewhat visible, but unfortunately, we can't really move this in a way that allows us to see more. Then there is also just a leather strap. So that's actually I will be honest that I'm more

25:07 impressed with this than I'd anticipated. It's not always a given that something even loads here. Look at the reflection on the face of the watch and we can see a bit more detail. Maybe

25:15 I'll just wait a second for it to make a full rotation. Some of the numeral markers are arranged a bit curiously. However, this is actually all right. Again, it is a test

25:26 that's been around for a little bit of time, so it is possible things go in the training data. But hey, look at that. We even have a visible portion of another one with a metal band

25:36 and actually like a decent front end, I guess, given that more of the emphasis of like focus on this is just how the model is. So that was our Q4KM result. So here is our Q8 result.

25:50 All right, this how can we do Okay, we basically just have to wait a little bit of time for it to make a full rotation around. On first glance, I'm going to notice the watch band is definitely more

25:59 in line with the size one would expect to be. So, there's a bit more realism there, but it is overall darker, so perhaps a bit more difficult to actually make out some of the finite detail in

26:08 the model. All right, let's scroll down and hope that we get a better look at some of these watches. I think overall, actually, this is a tougher one to judge because yeah, the cards here where

26:19 they're both moving around is nice. However, if we look back at this one, if you had made me guess which was which, I think I would have had a pretty difficult time actually guessing which

26:29 one was the Q8 because this could definitely go either way. So, it's interesting to see. And again, this is a task that maybe has been around for a while, so it's possible that models can

26:40 expect something akin to this, but there's definitely less of a visual striking difference between the two here. With a lot of the results, it's pretty clear like, okay, that's the Q8.

26:49 In this case, not so much. So, for our final test, we're going to be running something that I have actually been working on here with a Quen 3.827B. I've been doing a much more intricate

26:59 test with that model, and that will come out in the next few days or so. Hopefully, it's been a multi-day affair. So, in the meantime, I am giving these the same exact tests. Well, one of them

27:08 that I have been running with that model, which is to using Blender and GDAU create street. So, the traditional Street Eat game that we like to see, but they create the assets in Blender and

27:18 then they create the game in GDO. and then we ultimately play it. I will say without doing too much in the way of spoilers, Quen 3.827B did a rather impressive job in this specific task.

27:30 So, we'll see how these do. This is one I'm going to be running one at a time just because I want them each to have like free access to these tools. So, the Q4KM Street game seems to be finished.

27:41 And this is again using Blender for game assets and then GDAU for the actual game. So, as soon as I figure out exactly how to actually play this, we'll take a look at it.

27:53 All right, we're getting a bunch of errors and unfortunately, well, we can't really seem to move, but it did actually use both Blender and GDAU to make a game. It has a varying

28:08 level of completeness and honestly I didn't expect like a fully functional game from this. But the fact that it actually made something that does open and show us at least some semblance of

28:16 what we would have expected to see in a Street Eat game is good enough for me. And now for our final test. We're going to be giving the same Street Eat prompt just to the Q8 version of Ornith 35B.

28:27 All right, so we have our Q8 Street Eat result. That is actually really quite all right. It's totally a significant increase over what we saw with the Q4. The balding um

28:41 people's people are quite something. This is actually I'm going to say this is really really really well done. It made these assets. It put it all together. The heating does work albeit

28:54 maybe a bit simple and like not super exciting to watch. We don't have any like other mousees camera but every time I click it just ends up yeeting. So I can't actually

29:11 like move the view we have here. But it definitely works. I mean this is like this is really this is well done. The sounds the shake of the heat and the fact that it compacted its contact so

29:27 many so many times. This is really I think I'm pretty impressed with this and I would imagine folks will also be definitely a large large discrepancy between the Q4 and the Q8 here. And

29:40 really that leads me that's going to lead me into the summary here. I've been filming this over a couple of days. These are pretty good I think. I mean I don't know because I don't recall

29:51 testing the model these are based on which was Quen 3.5 or 3.6 6 the 35B A3B Quen mixture of experts model. I don't recall to have this depth of capability. I may in fact

30:04 be wrong, but I really think that folks are going to be overall fairly impressed with at least some of the performance in these tests here. In specific, this is one that has never ever been done

30:13 before. So, this isn't something that, oh, it made its way into the training data. So, that's why the model performed well in it. This is entirely a test that I've never actually shown on camera

30:22 until this video. So, this is pretty darn exciting. And I would say again the capability was pretty good. In some circumstances, not all of them, basically like with the watch website

30:32 and also with the 3D engine model, the Q4KM may arguably have done a better like greater than or equal to what the Q8 did. But in jobs like this or in the C++ skate game, we definitely see where

30:45 the Q8 kind of tends to pull ahead. So just interesting. And I wanted to do a side byside because although I like testing on Q8 because it shows us more potential performance, I realized that

30:56 Q4KM is something that a lot of folks are going to be opting for because it's more realistic to run. So that is why I wanted to do kind of like a side by side. And overall, that is probably

31:06 going to conclude our first look and test of the Ornith 1.535B- A3B models. They really are, I would say, just based off of this day or two of testing, pretty impressive. And I

31:17 think folks may find it worthwhile to try these. So with that, I suppose we'll do one more yeet if we can get the mesh collider to trigger

31:29 wherever it needs to. Good. And with that, if you have any questions, please feel free to leave them in the comments. And thanks for watching.