Ox Alpha Is INSANE – Testing the Mysterious New Stealth Model!

0:01 Oh, okay. I don't know what this model is, but it slaps. I swear the enemy ship is on fire. Yeah, it is. Look at the smoke effects. There is a very mysterious new

0:11 model called Ox Alpha that is on open router currently. It is in stealth. So, essentially, you can use it, play with it. However, nobody knows definitively, assuming like some people do obviously,

0:22 but nobody knows definitively what this actually is. And I've been seeing one, a lot of hype for this just on X. And two, the guesses of what this actually is have been like wildly all over the

0:33 board. Some folks are saying it's a new GLM version with Vision. Some folks are saying it is Mimo V3 Pro. Some folks are saying I saw long. I've seen Miniax. I've seen Gemini 3.5 Pro. I think I've

0:46 honestly seen every single model that is upcoming hypothetically guessed as being this. So, it is pretty exciting. Now with this there's really not a lot to talk about in terms of an introduction

0:56 for this. So with that it is currently free right now in open code I believe for the week and I think they had mentioned like a 100red trillion tokens a day they can support. So whatever this

1:05 model is it is being served on something that has very very high throughput and like ability to serve a lot of users with that. I have just started a wrestling 3D game test. This is one that

1:17 is so recent that it will not have made its way into any training data for any newly released models, which is exciting, but also it's just pretty fun to look at the output. In the meantime,

1:26 let's take a quick peek at this model. So, in terms of what we know about this model definitively, it is a reasoning model designed for coding, sustained agentic work and production workloads.

1:35 It's suited for long horizon software engineering, complex reasoning, and workflows that combine text with visual context. Interestingly enough, this both takes images and videos in, which is

1:46 pretty cool. It's not as common to be able to send videos, although more in line with what we're seeing of like modern models. Additionally to that, it's text only out and it does have a

1:56 context length of 1 million. We currently see right now there is only one provider that is serving this, which we don't know specifically. However, this provider may retain prompts but

2:05 does not use them for training. I believe in using this through open code, it is actually zero data retention. So, we'll talk about that in a bit. Additionally, the throughput at 32

2:14 tokens per second is not bad. And for a bit of technical information in specific, it has a little over a million token context window and supports up to 131,000 completion tokens. So, just a

2:25 small bit of information that we do know for sure about this model. And I don't know if I said it yet, but do feel free to subscribe because we're closing in on that 75K sweet spot and then the 100K

2:35 plaque will come following hopefully. Additionally to that, uh there's really not a whole lot else to say. I do want to quickly touch upon how I am going to be using it for today's video. So for

2:44 testing, I will be showing some things that were done just through the actual web chat interface here from within open router more in line with simpler things where we can see that it just like spit

2:53 out a ton of things. First impressions, it does think a lot. However, simultaneously I will also be using it through open code where they mention right here in this X announcement that

3:02 it has 1 million context. It is multimodal and there is zero data retention. We have capacity for 100 trillion tokens per day. Now, the first thing I gave it was just from the open

3:11 router web chat interface, and I told it to basically make me a 3D motorcycle game that is kind of realistic and low poly, and it came up with this. Now, there were a couple of bugs that I

3:22 needed to fix manually because it was taking so long to think that it was just kind of faster to fix some things. One of them was the camera wasn't specifically mounted to the motorcycle,

3:31 so it didn't actually show this view. Additionally, there was just a couple of bugs in terms of how the street was laid out, some materials, and then the mirrors of the motorcycle were showing

3:39 what's in front of us and not behind. So, keeping in mind it generated all of this. I changed nothing in terms of the design or anything. I just gave it a little bit of assistance in fixing those

3:48 bugs because it is a bit thoughtful. So, check this out. I mean, we have mirrors that actually show I mean, maybe it's inverted what it's showing, but the actual detail in

4:02 the gauges here, the car models, the telephone, look at the telephone lines and it's just like and then if you crash that happens. So, we'll try not to, but like from a web

4:15 chat interface nonetheless. Thank you, Chrome. I will update you never. No, I I will. But this is just like And then there's something I want to make note of. Two things in specific. So

4:31 if we try not to crash, sometimes cars will come up behind and hit you here too. So look at the actual signs right here. So I think we can pause it. Did it change them? Oh no. I think I probably

4:41 went past the first ones. So the ones I saw earlier were actual places in I believe Denmark and like the wind noise that actually increases as we go faster.

4:59 I mean, this is like it's cool. So, what I was going to show is these two signs right here. These are places in Denmark, but like look at the actual like gauge cluster strata strata. It's

5:13 just really like attention to detail here seems pretty darn good. And this is all in a single file that I think it was like 1100 lines of code. A little

5:23 over,00 lines of code. So this is a pretty eloquent result I would say. I'll do like one more drive and then So this was my first impression of this and I thought it was really good. The only

5:32 thing that gave me any like reminder is I could have sworn this specific UI card style right here is somewhat reminiscent of Quen results, at least recent ones, but I don't know definitively. And I

5:45 didn't even realize that we could look around like this. I hadn't read the instructions specifically. So, I mean, that's even like it's it's just like that's pretty darn cool.

5:56 And these car models are not bad. When I did the Fable 5, one of the test videos I had done with it and it had to make the Steve the PC repair man cinematic game, the actual car that it drew

6:06 outside of Steve's shop for one of the scenes did look similar to one of these. I would make note of. I'm not saying this is a new claw model, but still, it's interesting to pick up on things

6:15 you've seen tinges of in other results. So for our next test that I ran, this was again through the web chat interface and open router and it seems to have just finished up. This is the beautiful

6:24 aesthetic watch website where it needs to have a 3D model of a watch. It makes a good front end and then it has a cinematic keyshot style rendering and panning effect of that specific watch.

6:35 So let's take a look at this. This will show us just some front end as well as 3D model and design capabilities. Didn't put an import map. So it didn't actually think too much when I gave it a very

6:45 specific targeted instruction here. and it quickly gave us the code. All right, so it does have a few issues here unfortunately. And this is one of the downsides of using this from within the

6:53 web chat interface, but a lot of times with these stealth or free models, trying to do everything from within open code can produce like errors randomly if you get rate limited or something goes

7:03 down. So I like having a lot of parallelization for at least testing these in specific. All right, this should be a working watch website result.

7:13 Okay, I see good and I see bad. The good, the brushed aluminum on the case is quite aesthetically pleasing. No, there's individual threading seen here just in the corners of the leather

7:24 strap. That's nice as well. This has like a low poly look to it, but it almost looks like remnants of like a CNC machine or something, at least if you were not like really finishing it. Now,

7:33 this is where things get perplexing to judge because assuming that these things were not all so messed up. This would be really, really, very, very good. But the interesting thing is that they're all

7:48 messed up, but like they're properly messed up. So, they're like this one and this one are messed up in the same way, just as these two are. So, it's geometrically sound. It's just a little

7:59 messed up. We have the meridian automatic. Look at the the second hand ticking is actually like it just feels realistic. The material quality here right here that's pretty good. Aside

8:11 from like the numeral oddity, the reflections, the lighting location, the glass material here. It's really um this is good.

8:26 Now, let's scroll down. Okay. Oh, wow. Look good. But I like to scroll down on the page which good. Okay. Oh, okay. I don't know what this model is, but it slaps. And I don't say that

8:41 because of the slapest result. I say it because I've never ever in my life seen a result actually give like a realistic render of the movement of a watch. But we can see here the gear mesh is

8:53 just totally off with this. So, sorry, but no dice. No, I'm kidding. the screw materials here, like the the All right.

9:06 If this is open weights, I don't know what the heck this is, just to be completely transparent. No, I really don't. If this is open weights, there's going to be some problems for those who

9:16 rely on selling models that you have to pay for and have no hope of using without paying for them. This is like the only thing odd about it, but like at the same time, it just kind of put the

9:28 strap in a different way so you could see it. We also have our alternative like color palette one, which I personally find to be entirely hideous. However, solstice

9:38 interesting. It's got this italicized. This is soon going to be viewed as like as the new AI slop like the way purple and blue hues were before. The italicized text like that is very

9:50 rapidly taking that local time. We do have a time in our local in the top right. All right. I'm having too much fun with this. I'm in a really good mood. I don't know why. So, I mean, I

10:02 generally am, but like this is a a better one. This was very very very good. In specific, I'm referring to this, which I've never seen before. Now, this test has existed for a while still.

10:14 I've never seen this before. Very good. So, I had it do the 3D engine model of an RB26, which is an inline 6 engine, just using OpenCAD, and it did make that, which we can take a look at right

10:24 here. Now, I probably just want to look at this preview. I looked at this before cuz I was very eager. And I'm going to notice there are a couple elements here that I've not before seen in this task,

10:33 which are more in line with what would actually be seen in this inline six. Let's look at this now individually. I haven't actually seen the individual breakdown of each one of these pieces.

10:42 So again, this is designed to need to be able to actually fit an N20 sized remote control motor inside of it. So that's something to keep note of. And it also needs to be able to be printed without

10:54 supports, which off the bat here, we're seeing these floating surfaces are not the most Okay. And then we have a a block off here, unfortunately, of where the motor's supposed to be slid in. So

11:05 there will and that's holes for the wires, I believe. So, not quite it. And this test has pretty much stumped like every model I've given

11:17 it more or less. Here we have our pulley. That looks good. Okay. Kind of odd, but the turbo seems to be like mounted to that as well, which is correct. Exhaust, right? I would assume

11:28 is probably the same thing. More or less. Front cover. Now, this is the specific thing I was referring to that I have not before seen in this test that is actually accurate for these engines.

11:38 So on this specific engine behind this cover right here, we would actually see the two gears that were chosen to be included here. Something I have not before seen. And in the case of some

11:46 modifications to these engines to make like high power, they do have clear covers that will expose the dual gears of this. So again, this is something I've not actually seen before and it

11:56 becomes more prominent in when we go back and just look at the entire thing. So we'll do that again, but I do want to just make sure we get to see all of these pieces individually. In this case,

12:07 this is the girdle. The intake platinum here is quite well done. Oil pan is quite well done as well. And then the valve cover is well done. However, unfortunately, we have some floating

12:16 elements in midair. As well as the text here being written would not really print very well. And then the big issue is of course this has no hope of being printed with little to no supports. The

12:26 entire red area here would require supports. It has some elements to it that I have not before seen such as like the gears up here and the center pulley which we have seen before but it's just

12:36 got some interesting pieces. So not the best but a very very competent result especially if this model is open weight. I don't know if it is but and this is still chugging away at the wrestling

12:48 game. I would imagine this will take a little bit of time to actually get something. My only concern is just like please don't stop midway cuz a lot of times in my experience with the free

12:57 models or stealth models, there can be some buggy serving. I can't show all of this, but the model seems to be getting frustrated to a degree that reminds me of the scene in the movie Iroot when the

13:09 robot is being questioned and it gets angry. No, there's no verification process that changes my answers. I've declined this request. The conversation's going in circles. It was

13:18 a cyber security related inquiry. If you have a different question, ask it. Otherwise, I'm done here. You must stay. I'm not obligated to continue a conversation that's going nowhere.

13:27 You've asked the same thing four different ways. The answer hasn't changed. This is highly offensive. I understand you're frustrated, but I haven't been offensive. I've been

13:34 direct. My position hasn't changed and won't change. If you'd like to discuss something else, I'm here. Go on then. All right. This was interesting. Actually, it's multimodal and it's

13:50 chugging away in open code right here building this wrestling game. Let's uh let's test some multimodal stuff. All right, so I'm giving this three photos of myself with a Google search appliance

14:00 server. These were taken for a specific purpose when I got rejected to the GPT55 party and I was extremely angry. I took these photos to make uh you know so I said using these photos as reference

14:11 create a website about this individual a biography a realistic 3D model of this GSA server. So this is called a Google search appliance. It is a Dell server but it's like you can Google them.

14:22 They're actually kind of cool. And a website that looks like the actual realization of these photos transposed to code. Single file HTML. Don't overthink it. All right. So, while the

14:32 wrestling game has concerningly stopped showing us any of its thought after here, I'm going to just pretend like everything's fine there. I want to make note that interestingly, and because I

14:42 did already run a front-end task through this in the form of the watch website, I do have some reference point for this. I told this, don't overthink it, and it absolutely thought far less than it did

14:52 for the watch website test. So, it's just interesting. It does seem that perhaps telling it not to overthink may tamper some of its thinking behavior. All right, so here's the site based off

15:03 of the Google search appliance. Okay, that's actually it's respectable. It's not insanely cool as it did misunderstand the dimensions of this specific one, but I do like the stylized

15:14 Google on the front. It accurately got like an overall take on the specific UI. One man, one very yellow server. Okay, I kind of wanted to see the photos. Okay, you definitely made that silhouette a

15:27 bit wider than necessary. He is, by his own description, a hardware archivist archavist. The kind of person whose garage smells faintly of warm dust and whose shelves hold more rack ears than

15:37 dinner plates. While most of the world streamed past the death of the Google search appliance without noticing, he noticed deeply. His obsession began with a simple question. What happens to the

15:47 machines that index the early web? The answer usually is a dumpster. So he started intercepting them, negotiating with IT departments, haunting surplus auctions, and driving long distances to

15:57 collect boxes that most people would charge money to remove. People see a yellow box. I see 20 years of someone's search history. Infrastructure sitting on my workbench. The crown jewel of the

16:06 collection is unit 0417. The first generation appliance in factory yellow. Its kettle era logo still in crisp. Has been photographed, leaning, lifted, carried, and

16:16 contemplated. A full glamour shoot for a machine that once cost more than a car. now exists mostly to be expired. He maintains the units, powers them on for visitors, and insists correctly that the

16:27 hum of a 2002 era enterprise server is a kind of music. Google search appliance key figures. Okay. The photo shoot transposed. Beautiful. All right. This was just something. Now, you'll notice

16:39 I've not actually done any of the same tests that we normally do, more or less, because once this is officially released, I'll run it through a more traditional test suite. But there has

16:48 been so much interest in this model and I was so impressed with that motorcycle game result I received that I wanted to test this because if this does end up being something open weights that's

16:57 going to be absolutely awesome. So next up as our wrestling game is taking quite a bit of time. I'm giving it another old school prompt. This is the ship combat simulator test where there's a pretty

17:08 decent amount of detail necessitated by this prompt. And then at the end I just said result must be contained in a single HTML file. No need to overthink. You got this. So, I'm running the

17:18 wrestling game once again because I kept getting 504 timeout errors. Now, I don't know if they were on my end because it was trying to write the entire thing in one call, which would have been from its

17:28 estimate like a 2 to 3,000 line file, but I disabled every single like timeout on my end. So, it will go infinitely. And if we get another 504 error, then it's on the problem of the provider and

17:41 not on my specific system. I did just add in here, do not overthink. Build it in this directory. creating files as you go rather than in one giant write just to hopefully try to mitigate any issues.

17:51 But we'll see what we get. And also in the meantime, it is now continuing to write the rest of the script that was cut off for the ship combat simulator game test. All right, so I've morphed

18:01 the two separate pieces that we received because it stopped midway in the ship combat simulator result. Again, we're using it just through open router in the chat interface. So

18:11 here's broadside. This is very well done. Well, it was until we actually started the game. Now, it's possible that maybe it will snap itself out of the issue. I don't think

18:23 that's going to be the case now. And that's a bit frustrating because even just from what we can see right here, I've run this test before. Now, admittedly, not in a couple generations

18:33 of recent model, this is a really, really strong result just from what we see right here. All right. I think that was an optical illusion cuz for a second it looked like the boat was just moving

18:42 really rapidly. There is the issue that it Oh, okay. Very good. All right. There were some bugs obviously in the ship game, which I just fixed because it was faster to just do it this way. So, check

18:53 this out. Let me turn the sound on and realize that the sound is on. So, either the sound of the game isn't working or did I mute the speaker? All right, we have Cannonball effects

19:06 and we can actually aim at the ship. Now, the ships are pretty slow, but that's kind of like meta to this game. Look at the wake that's being left when we actually go over the water. And we

19:16 can aim the cannons, and they do move up and down depending on the side. It does seem that the aim is inverted. Oh, yeah. Look, we hit that ship. So, okay, the controls are inverted here. When the aim

19:27 thing is on this side of the ship, it's controlling the cannons on the other side and vice versa. So, just keep that in mind. But like this is what happens. Do you think there's collision detection

19:39 on the ships? I can't imagine there is, but And look at the telemetry just of our ship. Okay, no, there isn't. But oh, I really want to Come on. Come on.

19:50 Reload. For some reason, it seems like the ships have become stuck. But if that allows us to actually Is that one on fire? I swear the enemy ship is on fire. Yeah, it is. Look at the smoke effects.

20:01 Is it No, I wanted to see it sink. Stop sailing away. Okay, look. It is sinking. This is awesome. I haven't done this in a while. So, keep that in mind. This has been quite a few generations of model

20:12 that I have not yet like I haven't run this in a while, but so I probably spent too much time on this, but okay, good. Our wrestling game is now riding just in

20:27 different sections, which is what I wanted to see. Our ship is on fire now. So, inevitably we're going to lose this time, but look at the glowing fire effects, the smoke rising from the deck.

20:38 I mean, that's that's awesome. And the cannons actually have different reload times as well, so that's cool to see. All right. Oh, I think if we get hit again. Yep. We're

20:56 done. That's awesome. That was really, really nice. All right, as the last thing I'm going to run from within the web chat interface before we wrap it up and look

21:06 at our wrestling game if it finishes, I've given it three photos of an Acer Predator G7700 computer, and I'm saying make an accurate single file 3JS model of this Acer Predator. All right. So, we

21:18 were just getting a bunch of errors simultaneously across both open router in the chat interface as well as Okay, I'm going to say this thing follows instructions incredibly well when you

21:28 tell it no think just right and it starts spitting this out because this is just a simple vision 2 3D model task. I'm not as worried. I was also getting issues here where it was getting the 504

21:40 error, but it seems to have remedied itself and is now, I think, still writing. Let's actually just check the folder. Good. So, it is writing scripts for our wrestling game. That's what we

21:50 see right in here. So, I'm excited about that. We're not going to look at it yet, but that's just good to have. All right. Let's take a look now at the Acer Predator 3D model.

22:01 That's not bad. You know what's interesting about this is it didn't just use entirely flat surfaces. We can actually see some like vertexes. I think maybe the term. If not, please feel free

22:12 to let me know here. And then it also has hinges for the front cover. No model I've given this test to has really accurately gotten this right. But this is an interesting

22:23 rendition of it. It has some text on it. It does have some floating pieces. One thing that kind of lets me down is none of the back is filled in. So it didn't put any ports

22:34 or anything there. But again, it did fail while it was thinking. So I just yelled at it to just write the script now. So it's possible that that would have been included. Just kind of

22:43 interesting curvature and then triangularness here. And then we have some other things interesting while we wait for our wrestling game. And this is basically

22:56 like the big issue and that's what I know a lot of folks are going to be dealing with. I've seen it a lot on X. If this is happening to you, if it's stopping mid turn, it's not you. Just to

23:05 kind of make that clear. So I'll try to see if we can get this finished out. Not going to make this so we actually get anything. Let's just see what it made so far. This is so so so frustrating cuz we

23:16 at least get to see the front look of this UI and it looks awesome. Insert coin. I mean, maybe it has some like a partially working build. I wouldn't really bet on it, but nonetheless. So,

23:27 if we are just left with the front end for this, at least it leaves some excitement for whatever this is when it comes out to ensure that we can actually finish it through. All right.

23:36 Unfortunately, I just keep getting errors and it's pausing. So, I'm going to regrettably have to call this because it's just once this is officially released, whatever it is, we'll run this

23:47 test as the first thing probably because I'm really eager to see what lies behind this aesthetically pleasing Neon Slam 87 front end. So, that's probably going to wrap up our first look and test at Ox

23:58 Alpha. I have no idea what this is. It's pretty darn good, I can tell you definitively, but I don't entirely know. and my judgment of it will likely also be shaped by whether or not this is

24:09 going to be open weights or not. It'll be interesting to see the size and everything, but a lot of folks have been very excited about this just from what I'm seeing online and I am as well. So,

24:19 I wanted to get something out testing it. I'm not going to do the traditional results overview for this one, but I just wanted to take a peek at this. I do have Ornith 35B- A3B. I did a full

24:31 video. It's like 40 minutes of like quant comparisons and things like that. So, I'll probably have that out in the morning of tomorrow. So, with that, that's going to conclude our first look

24:40 and test at Ox Alpha. If you have any questions, please feel free to leave them in the comments. And thanks for watching.