Qwen3.8 Flash Next Is INSANE – Testing the PREVIEW of Qwen 4!

0:00 There's even water. Look at this. Just even if we're looking at the terrain here, Alibaba has released Quen 3.8 Flash Next, which is a very exciting model

0:12 because it is a preview of the architecture that is going to underly the Quen 4 family of models. So, if you recall, there was a Quen 3 Next model and that was a preview of the

0:23 architecture that we've used for Quen 35, 3.6, 3.8, etc. And this is kind of like the next step in that. Get it? So before we get into it, do feel free to subscribe so we can hit that 75K number

0:36 and then the 100K plaque. And let's begin just by taking a look at this model and some of the interesting things about it. Now admittedly, there is a lot of kind of technical meat in this model

0:45 and in this architecture here that's likely worthy of its own like portion of the video. However, truthfully, I think we're just more interested in testing it and seeing how it does with a subway

0:55 FPS. No. But with that, there are a couple of things of note that are interesting. And one of those is this right here. So this has an engram embedding where they mention it looks up

1:04 a table using the local context to scale model capacity with very little extra computation. So this model right here is a mixture of experts model and the specific number of parameters was

1:15 mentioned somewhere. Nice. Okay. So yes, it's 125 billion parameter main model with 6 billion active and then it has an additional 51b engram embedding. And the cool thing about that is what they

1:28 mention right here. The embedding cable can be offloaded to host memory and overlapped with model computation through async prefetching. So basically this is something that enables increased

1:39 capability in the model that is actually okay to be offloaded to CPU RAM in lie of assuming all these numbers were stacked up. So you would think like oh I have no hope of being able to run 176

1:52 billion parameter mixture of experts model but instead this is specifically okay to be offloaded to CPU RAM. So it lightens the overall weight of what's needed at least for like a big beefy

2:03 GPU, but still increases capability. So I mean you would still need the like amount of system RAM as well, but it's just a very interesting architecture and it's as they say the next model. So with

2:16 that out of the way, our next step get it. No, let's just take a quick peek at some other interesting things about this model. Now currently this is open weight on hugging face. There was a nice

2:26 countdown for it. And while I personally can't run this locally at a higher precision quant, they do now have a Unsloth, who did have Day Zero support for this, which is pretty cool. But they

2:37 have released their 4-bit quant and the Q4K XL is 111 gigs. I'm currently downloading that on the system behind me. And while for today's video, we're likely going to be performing most if

2:49 not all of these tests just through the API so we can see this at max performance, I am downloading this and do have every intention of at least trying to run it to see how the speed is

2:58 and everything like that with a lot of CPU RAM offload. Additionally to that pricing, so if you can't run this locally, it's priced at 16 cents per million in and 47 cents per million out.

3:09 That's a pretty cheap price though, of course, we want to see how this thing actually performs. And they do have some benchmarks here. And in specific, I'm very interested in seeing this against

3:18 V4 Flash. However, unfortunately, the new version of that, which has experimental vision, which is really quite potent, is not included in this benchmark. So, keep that in mind when

3:27 we're looking at some comparisons of the scores for this. I'm really interested in this one right here, the Deep S. SE. And that score right there is really quite decent. I do believe the V4 flash

3:37 model with vision has jumped into the 60s for this specific benchmark. However, that is a bigger model by a decent amount. So, this still seems like price to performance or weight to

3:49 performance is very potent. So, for some additional information right here, we can just see it's multimodal with video and image in which is good to have. Max output is 131 and then max input here.

3:59 So, max context for this as served on Quen Cloud is a million. Keep in mind just on Hugging Face right here, it's going to be 262 144 at least natively. Now, in terms of how we're going to be

4:10 testing this model today, I am just using it through some middle tier token plan they have. I did just resign up for it. If you recall in one of my previous Quen videos, I had assumed I was using

4:20 the token plan and I was just using non-subsidized billing. So, at least it gave us an accurate assessment of the cost of testing. I did just subscribe for this myself. Uh, it had lapsed. So,

4:31 and additionally to that, I am currently downloading the Q4K XL Unsloth Quant for the system behind me. I don't know that this will make its way into the video, but I would like to at least see the

4:42 comparison between how a local 4-bit quant gives us results versus the cloud version, which is going to be running at max performance. And yeah, I also want to just see how it does with some

4:53 offloading and some speed measurements and things of that sort. So, with that, I do have this running currently in Pi. And we're going to begin, of course, with the triedand-true browser OS test.

5:02 However, I have given it a slightly different task, which I will showcase right here. I've given it this reference photo because I want to do some more multimodal testing with these models. I

5:11 think it's kind of fun. And I've said I would like a pixel perfect replication of this image, but the rest of it is the traditional browser OS test. So, it needs to have a GTA clone, procedural

5:21 wallpaper, and email client, etc. And we'll take a peek at how it's doing now. All right. So after quite a bit of time and keep in mind this is a preview model. So

5:35 we can assess the vision capabilities are nothing short of incredible. Is this all right? That is by far the clearest replication of this that I've seen

5:48 period. What I was going to say before I got stopped was this is an experimental model. So it may not be 100% like all there. However, this is like actually let me bring up the reference photo. It

6:06 moved it into a Oh my goodness. Look at all of this work it did. We have heat maps. We have I think it sliced this into specific just based off what I'm seeing here. I don't know. So don't

6:18 quote me on this. I think it's I think it sliced this into individual pieces and then like reverse engineered them one by one kind of like what we see right here. The the work that's gone

6:34 into this is extremely extremely impressive. This is the reference photo and then this is what it made. That's incredibly well done. Really incredibly The Mac OS icon is blinking. I've not

6:49 seen that before. Now I can't move these around easily at least. So, um, that's a failure. No. All right, let's just let's look through this. 78* 6 468.

7:05 Good. All right, let's uh All right, we have the trash. I can't click that, but that's okay. Uhoh, hold on. No, no. New founder

7:19 window documents applications. Yes. All right. So, we have our apps here. Let's just try the GTA clone.

7:34 Okay. It not not quite street. However, there's like some walking effect. And again, this is a preview model, so it's going to improve. And this is just

7:50 an early look at the architecture of Quen 4, which just based off of its multimodal cap downhill snowboard. What? Finally, they didn't put a darn space shooter game. All right, this is like

8:08 I don't know why I made that noise. Look at the way it did it, too. I don't have much to say. It's just internet mail. Look at this. We get an email from like they're meta emails

8:26 here. So, like about the games and things like that. Note to self, buy more floppy discs. Lot of creative writing gets shown in the email client for the browser OS tests,

8:38 which is an unexpected side effect. So again, this was a suggested improvement to this task from one of the comments. So thank you. My pad calculator we saw. What is this one? Appearance.

8:52 Oh, is it moving? Yeah, it is. I think solar gradient water alien bridge line starfield. And look, it's actually like these don't look modern. These look in the same era that this would be.

9:07 Especially this one right here. Speed. Yep. Screen saver start after blank seconds idle. This is actually like test now. This is exceptional.

9:23 Is there a special feature? Even if there isn't like I don't doesn't need to be. Time machine. Yeah, that's what it'll be. Snapshots go to now. Oh my.

9:39 All right. This was this was like this was top top top tier. It even did it in like a non widescreen look. So it really like pixel perfect replicated this shutdown.

10:00 Now, in the meantime, Pico Kumbbury, I had given this a task to make a C++ 3D lowpoly rally game. I didn't want to give it the skate game just due to the fact that it's so long in the tooth. It

10:16 could be benchmaxed. I wanted to give it something that there's really no prior reference for. I've done this a few random times. This is really, really, really, really good because it was going

10:27 to use Ray Lib. I stopped. Jeez, I stopped it and said don't use Ray Lip. Our rally game, I think, is pretty much basically done as it was talking about just the final cleanup and summary and

10:40 doing the read me. I mean, this is really like it's maybe not fully finished yet, but and I love that you can actually drift and like crash and things like this.

10:55 really as a C++ result for a model of this size. It's incredibly well done. At least for whatever was Can we do like donuts? Not fully. And there are even folks in

11:08 the audience. Well, you know what I mean? Like in the stands. And I think really the big takeaway from this is even if these results are not altogether right now, the fact that this

11:20 is a preview model of the next generation should be very very bullish and we've kind of crashed. We're going the wrong way.

11:31 Even the look at the terrain and I really mean it. This is a result that would have been this would have eclipsed a state-of-the-art model from a year ago in terms of this specific task without

11:41 being able to use ray. And I know that because I've sparsely done this in the past, not as often as the skate game though, but like there's even water. Look at this. Just even if we're looking

11:50 at the terrain here, it's like there's almost suspensionness to the car. And when we go off the map, it doesn't restart us. It actually This is really like the

12:04 I mean Oh, whoops. Okay, so this must be the uh because it was still wrapping it up. This must be like the wrapped up version more or less. This is like actually fun. I should probably be

12:21 continuing this right now, but it's so enjoyable to play this. It's a little funky once we get through this section. The car is bucking like that because it's hitting the rev limiter in first

12:34 gear and it's not shifting, I think. Yeah. And we can actually like there's a E bra. This is nuts. All right. I mean,

12:46 it's probably even the menu. Look at this. Next up, we're going to be giving this the 3D wrestling game, which needs to have a 1980s aesthetic. In this prompt, nothing says it needs to be an

12:57 HTML file or any specific language. Now, I've noticed here it immediately started to look to see if there was GDAU and Blender on this system. It's planned out and decided that it's going to create

13:08 the assets in Blender and then the game in GDAU. I don't say that it needs to do that, but in the same time, I don't say it needs to be contained in a single HTML file. So, I am a little concerned

13:17 at how long this may take or how much work it may end up doing for this. But, we can see it's been chugging along now for a little bit of time and it seems to be creating some things. I really have

13:27 to stop myself right now from peeking in some of the asset folders that it's created to just peek at some of the assets. I'm going to because to be honest with you, this was so insane that

13:37 I want to see what it's created so far just for the assets. This is genuinely I can't show this in the video. This is This is absolutely terrifying.

14:13 If I had if I had not looked at that and just waited for like however many hours this thing's going to take to finish this and seen that I would have probably passed away on camera from laughing. So

14:27 right now the big block the big blocker that we're having is literally this gigantic white blob and it's trying to just figure out exactly what's causing this. And it's a

14:40 little frustrating because like the game's actually kind of working. Like we can see a bunch of elements. Look at this. Look at the chair models. And then it even has a belt there. So that must

14:49 be like the championship belt. There's stats for the other wrestlers. Okay, we definitely just got yeated out of the ring. The biggest problem is whatever the heck this is. So it's currently

14:59 trying to figure out exactly whatever the heck this is. And we can see contextwise and things. This is very very very verbose. But still like and this made these all. So, this is using

15:09 GDAU and Blender for the assets as we saw. I still haven't seen the wrestlers. That's part of the problem with this blurring out. Maybe this is good. I might have had to do this anyway. I have

15:23 too much fun with this. I don't know why. I might have had to do this anyway, depending on how the wrestlers look. So, I'm going to just just going to close this for now. All right. So, after a bit

15:32 of time, Quen finally figured out how to fix the white cloud that was in front of this screen. It was like still I think like trying to debug it, but to be honest with you, it's done enough that

15:44 it seems like we're in a good spot to actually take a peek at this. So, I'm going to just ensure the sound is up. Okay, perhaps music is not its strong point. Let's look at our Wow, look at

16:00 the wrestler models. Enter space. Lock it in. Oh, what is wrong with Mr. Steel? I didn't realize that. Uh, maybe

16:17 let's just go with El Scorpion. All right. And keep in mind, it did this using Blender and GDAU. This is not 3JS. All right. Uh,

16:29 so they don't they unfortunately something's not 100% right, but it says down, it's frustrating because

16:44 it's the potential here is vast. And the fact that this is a preview of the next generation architecture and it did all of this using Blender and GDAU, not just opting for 3JS,

16:56 I still think this is a win even if it's a little troubled. Now they're it stinks because it seems like they fall through the floor when the ring actually starts. So Oh,

17:10 these games are all right. I think this is the most we're going to get from this, at least for the time being. But I do intend to perhaps do some more with something like this. So, this is not the

17:22 last that we'll have seen of Steel City Rumble 86. Out of curiosity, I also wanted to go back into the preview folder where Quen had actually made renders of the things it created in

17:32 Blender. Now, yes, these are odd, but if we judge this, I mean, these are in poses for fighting. It has an entire like the referee. Look at this. This is really quite

17:46 impressive. The roster right here, this is really for it's just showing some capability that I'm impressed by, I guess, should be the takeaway from that. And the fact

17:58 the amount of work this thing does is really quite impressive. Yes, these look extremely terrifying and like kind of scary, but it really went hard in actually like trying to properly emulate

18:10 realistic looking people with muscles and the outfits they'd be wearing, the poses, even putting them in these fight poses. This is really the amount of work that went into this is

18:22 incredible. So, we're now doing a true local test with this at a Q4KM quantization running on the machine behind us. This is the unsloth quant and we've given it the subway FPS test. I

18:33 want to give it something relatively simple to accomplish just to see how this performs in this quant. As this is truly the first time I've actually tested it for real local, so there could

18:42 be a mismatch, but I have ensured that everything is set up to spec as denoted in the hugging face model card just with the preserved thinking, all the sampling parameters and everything like that. And

18:51 it is on the highest selectable reasoning mode. And we see the difference here is the context is capped at that 262k as opposed to using it through clen cloud which we're doing in

19:01 the other terminals. It is set to a million. So, but everything just on first glance right here looks more or less competent. So, it's really cool to see this running locally. And again,

19:10 this is at a pretty decent speed. The specific specifications of the machine this is running on. RTX Pro 6000 and a Thread Ripper with 128 gigs of DDR5 that I just put in it probably like eight

19:24 hours ago overnight as I did want to add some RAM to that machine. It was running on a 32 gigs of DDR4 motherboard with like an old i7. So that was kind of a mismatch for that card. So that's the

19:36 speed that we're getting here. I don't know specifically what token speed, but decent. All right. So, the system had crashed because the process that it was doing right here for the Subway FPS just

19:47 from the local machine behind us. For some reason, it was ballooning in memory. So, the session it was running in Chrome to try to troubleshoot this took like 12 gigs of RAM and then the

19:56 system froze up and it died. But, we did get to a point where it says no errors, zero leaks, AI path climbing works. Now, testing the combat loop at close range. Because of that, I think we're safe to

20:07 basically take a peek at this. Now, I saw kind of a spoiler for it just preliminarily when it was loading in. And again, I would like to mention as well two things. One, this was done with

20:16 the Q4KM onslaught quantization. So, the UDQ4KM and this is a test that's been around for quite a while. So, if it looks really good, it may be because this is

20:28 in the training data possibly, but like this is pretty darn good so far. We do have the flashlight. Okay. Oh, I thought this pile of trash was an enemy. Let's just look around. Okay, we have a subway

20:42 car. We have time. Null. We have graffiti. I don't see any enemies. And it's a bit laggy. The weapon model looks all right.

20:58 I don't see Oh, okay. That was the most like wimpy sound I've ever heard for ammunition in this test ever. Oh, okay.

21:12 I don't know what just happened. Did I? Yeah. So, something's not like all there with this, but it was a 4bit quant in a preview version of the model. All right. Good. I mean, to be expected at this

21:26 point. Just the the ammo sound is so We're reloading. It's so wimpy. Is our flashlight on? Yeah. Oh, okay. This game's got Oh, did you see that? It

21:48 was like a health pack. This game's got the creeps big time. Like the environment and the enemies are moving backwards towards us. That's something that is

21:58 oddly fairly common. Oh, and maybe we lost and that's what's causing that. I don't 100% know, but I saw enough there that I wanted to do something from the actual local machine. So, we didn't just

22:11 run the entirety of this on the cloud. Next up, I'm going to be performing a test that we also did with the Deep Seek Flash experimental model that had vision where I'm giving it a few reference

22:21 photos of the room that I am in right now. and we're going to tell it to using Blender create an isometric scene of this room and then just show us the render, make sure everything's textured

22:31 nicely, etc. So, finally, we did a room render in an isometric scene. And a couple of times now, it's been like, okay, I'm going to write the final files, and it will have one more

22:39 improvement, and then it'll say, let me do the edits, and then it'll render them for quite a while. Then it will look at them and be like, okay, I really like this, but I think one thing I might want

22:47 to change is this. And then it kind of goes and does that again. Now, this survived because the system did crash. So, I had to restart it and resume this. So, keep that in mind as well. But I

22:57 resumed the Pi session, so it shouldn't be much of an issue. Let's just kind of take a look at where this is at. Photos or source photos. Okay, I should probably know that by now.

23:09 That's No, it's like did it. Okay, look what it's done here. It actually sliced up these source photos to use as textures for this

23:27 render because it would not have actually All right. Okay, I know what it did. See these paintings? It wouldn't have remade these. actually used these sliced photos

23:41 for the textures here in this scene. Even the purple lamp in the top left of me corner. That's not something that's really easy to see, I think, in these photos either. So, this is overall a

23:51 very troubled result. The pieces are not really put together at all properly. I've noticed consistently Quen models seem to have some form of issue with like spatial and things like that. But

24:01 if we judge this just based off of some of the individual elements, this could potentially evolve into a very, very nice incompetent scene, we see it's still chewing away behind us and

24:11 actually working. But if we look at some of these things, like if you squint at that and look, it's like, oh my goodness. Now, interestingly, the whole scene is kind of inverted. So, however

24:21 it went about reading these photos, it kind of got a bit confused because this PC is like on that wall. So, it should be like over there. It's almost like a weird way to try to put into words how

24:35 it's inverted this. And because of what we're seeing here, I feel a little more comfortable stopping this prematurely because look at the cables even that it's running onto the ground because

24:44 it's not quite understood the scene. So, again, for a preview model, I think this is quite potent. And I do want to just take a look at these in Blender. So, here's the Blender scene that it's

24:56 created. And we definitely see more of how it's troubled, but at the same time, it's also quite something. I don't quite know what to say. Definitely. I think Deepseek V4

25:09 Flash did a better job than this because even though the actual like detail of that one was not to this degree where it just basically used the existing textures from the photos that it

25:20 actually placed into this, it overall did a more competent job because it accurately reflected the scene better than this. But individual detail in this, even if you just look at it like

25:29 this for half a second, it's like, whoa, this thing's realistic. And then you zoom in and it starts to become more problematic. But definitely really exciting, I think, for a preview model.

25:40 And it's still just spinning out previews here. But because it's doing the same thing where they're not really all together, I think it's okay to kind of call that one. I think overall that's

25:48 probably going to conclude our first look and test of this model. I have every intention of doing some more tests, more intricate and different variety of like tasks with this GLM53 as

26:02 well as Deepseek Flash Vision Experimental, whatever it's called. So, this was really overall for a quick results overview, I'm going to say I'm most impressed with the vision

26:11 capability of this just based off of the replication of this source image for this screenshot. I've done this a few times and I've never seen a model do it this accurately and the large amount of

26:21 work that it actually put into doing this. So, when we go back into the reference folder here, it basically sliced up the source image. It did a heat map. It was doing side by sides to

26:31 properly see if it was actually able to replicate these things in a more pixel perfect manner. There was one, I can't seem to locate it right now, but with the two calculators side by side and

26:43 just overall the amount of work that went into this and it was clever as well. It had like some Easter eggs and the email app had some references to apps in this and just the way that it

26:53 really really did a nice job with this. I was taken back by the first look at this. Something else here that again really like absolutely came as a shock was this rally game. This it does have

27:06 sound. How does it you Okay. Well, now we know this was like the shifter, the interior, and just the overall like scenery. There was water. There was a bunch of other

27:25 things. We could drift. There were even folks watching the rally stage like that. The mini map, the speedometer, and like and it bucks like that. It won't like go

27:36 out of first gear. Okay, that one was really, really highly impressive. And I want to specifically mention why. So, here's a test of GPT 5.2 Pro. This is the Pro version, which

27:50 is always significantly significantly more performant. This is its result for the C++ racing game test. It may not be the exact same test, and this actually used RayB for this. The Quen model we

28:03 got a result for did not use ray which would make this more difficult. And look at the difference between the two of these. So if you're thinking like man you're overhyping it. It's not that

28:11 impressive considering that a 52 pro model did this with ray compared to what we saw with the quen model. Totally totally impressive at least for democratization of intelligence and

28:24 capability that can actually reasonably be run on local systems. So that's why I'm very excited about this and it's just something that's really cool. Yes, this test has appeared very sparsely in

28:34 some of my videos in the past, but I would bet that the capability would translate over to other unseen C++ tasks just based off of what was seen in Arwin result. And for specific reference, this

28:45 was on December 17th, 2025, so 8 months ago. Then we had our wrestling game, which was unfortunately not all there, but the way it went about actually making these assets from within Blender

28:58 and then creating the game within GDAU. From right here, this looks like a pretty darn impressive scene. And when we went over some of the things it did to actually create the renders of these

29:07 folks, like looking at these source images of the renders, it's pretty darn impressive that even if they look kind of scary in some regard, it did an interesting job of doing this. And

29:17 again, this prompt did not specifically say it needed to use Blender in GDAU. It looked at the prompt and decided to do that. A lot of times we end up with a 3JS result, which definitely would have

29:26 been significantly easier, but I'm glad it wasn't the result of this because this is a more impressive overall scene until we start playing it and then everything kind of falls apart. Then we

29:36 had our room render, which unfortunately just never really quite made it together competently, but when judged as some elements individually, basically the paintings and the Apple history poster,

29:47 it's like, wo, that's quite impressive. The way it just reused the textures from the actual image, I thought was pretty cool. And then there is some sign of capability here in some sparse elements

29:58 like the floating fans for the PC behind me. Even the color that has lightup DDR5 in it. and this actually kind of grasped that and tried to render it, even just the light on the CPU cooler. So,

30:10 interesting attention to detail like that, but unfortunately just not really competently put together. Then, finally, we had the subway FPS, which was kind of hilarious sounding because the weapon

30:21 sound was just very, very wimpy, which was funny, but overall not bad. And the coolest part about this was the fact that this was the unsloth 4-bit quant running locally on the system behind us.

30:30 So, it was nice to be able to do one of those as well and see that even though it crashed the system because it was using too much RAM when trying to play test this, it did recover and was able

30:39 to produce something. And again, this test has been around for a while. So, it's possible that this is a better reflection of the 4-bit quants capability than perhaps something that

30:48 has not been seen before. But overall, just cool to see. And I wanted to make sure to actually be able to test it locally, even if it was just for one thing like this. So with that, that is

30:58 probably going to conclude our first look and test of Quen 3 Next Flash, whatever the specific name for it is. I would say overall, as a precursor to the upcoming Quen 4, and this is overall a

31:09 preview of the architecture, I'm very, very, very excited. I think its vision capabilities are extremely potent. But again, I wonder if it does have some work to be done in terms of

31:19 the spatial orientation, which is something I've seen consistently as a weaker point in testing many different generations of Quen model. So overall, I think it's a pretty exciting day because

31:29 now it's time to test GLM 5.3 Flash and then whatever else happens to come out in the future. But the fact that so many competent models in around this size range have been coming out lately does

31:40 make me quite excited for the future of like nonsubscriptionbased AI. So with that, if you have any questions, please feel free to leave them in the comments.