Qwen3.8 MAX Preview Is HERE – Is THIS the BEST Open Model Yet?

[laughter] I didn't click stop. So, Alibaba has announced a new model, which is the Quen 3.8 series, and this is the first public mention of a Quen 3.8 model. Now, before we get into it, a couple of things. One, please do feel free to subscribe so I can get that 100K plaque, assuming you like my content. But two, this is pretty much going to serve as the entirety of our announcement post because this is currently still in preview. So everything we know about this model is contained within these few sentences right here. I would imagine in the coming days or weeks when this moves to a more general availability or whatever the terminology is. There will be a big announcement release post. There will be benchmarks and things of the sort. However, in lie of benchmark JPEGs, we do have a very, very bold claim right here saying that this is second only to Fable 5. Beyond that though, there's some other really exciting things here. One, this is going to be open weight. Now, historically, Alibaba Cloud, their Quen family of models, the Max ones were always closed source and proprietary. No previous Quen Max model has actually been openweight. And that's very exciting that they've decided they're going to openweight the Max model. It's I'm I was surprised to see this, but positively surprised I guess could be said. Additionally to that, the size this is 2.4 trillion parameters. That's massive. And previous Quen Max models were, I believe, somewhere in the neighborhood of around a trillion parameters. So 3.6 max, 3.7 max, etc. This is significantly larger than those. Though it is a bit smaller than Kimmy K3, which is 2.8 trillion parameters. it gets kind of like keeping track of all the different sizes gets a little cumbersome. [snorts] Now, something I want to touch upon here is that okay, this is open weight, but like no one's going to be able to run this locally at home. So, what is exciting about that? If it's open weight, but like no one can actually use it. It is true. Someone who has a system that is not five figures will have no hope of being able to run this. And even if you had a five figure plus system, you would either have to run this at such a extreme quant that the intelligence would be degraded to the point where it would not compete with Fable 5 or even much smaller models or it would be so slow that you'd get like half a token or a token per second using a big CPU and system RAM only server which I am in the process of planning out to build because I really do want to run some of these locally. But the exciting part is that this being open weight means a bunch of different providers can download these weights and serve them themselves from their own infrastructure from their own physical locations. And that is exciting as a consumer who pays and uses these models because when we use Anthropic or OpenAI, we use their model from them and we pay them directly. There's no alternative options except for maybe some niche like more business use cases where oh it's being served from bedrock or something like that. But for the vast majority of people, we have one option to use those models. When a big model like this that is performant is open weight, sure, it's not able to be run locally on hardware we own in our own home. However, there's a bunch of different providers that are going to be serving this. Some will have it being served faster. Some will have it being served cheaper. And it really opens up a lot of different options for folks who want to use this model as opposed to the flip side of that coin where you have one option and you pay this one price. So even though it's not necessarily able to be run locally, I sometimes see the sentiment expressed like why should I care about these models if they're so big that there's no hope of running them. And I just think that the democratization of them actually being served by more people, giving the consumer more choice is still exciting even if you can't use it at home. So with that, we don't really have much else to say about this. We know its size, we know its claimed performance, and we know it is currently in preview. So, if there are some oddities to it, hopefully they'll be ironed out by the time that this is no longer in preview. But, I'm not going to wait till this is officially released to do a video cuz I'm very, very excited to test this. And of course, we're going to begin just with the tried andrue browser OS test v2.5, which we can see has now begun to actually be written. This thought for quite a while. I do believe that I sent it this initial prompt. If we can scroll up all the way, probably 10 to 15 minutes ago, and it thought for quite a while. If I click on this, does it show us specifically how long it thought for? It doesn't, but look as I scroll up. I don't know if it's easy to see, but the scroll bar is there. It's very small and it's still very close to the bottom of the page. So, there was a large large amount of thought that had gone through this result. And now we just reach the top of that chain of thought. So, very very very intricate amount of thinking. All right. So, we have our browser OS test from the 2.4 trillion parameter Quen 3.8 Max preview. This is a very different UI splash screen than I've ever seen for this result ever. Uh it said the only external thing was like some font pack. I don't think that would be tripping it up. You you silly silly model. All right. So, I'm actually happy to see that it just gave us the patch because it shows some behavioral behavior in that like it knew like I'm not going to regenerate this entire script. Very good. Oh my. Art is subjective. Personally, I love it. I absolutely dig this UI aesthetic. It's just very, very, very different to anything I've seen before in doing this test. Seriously, they're kind of all the same at this point. Not this one. This is a very, very interesting style. Now, good. We can resize. I love this. I really like this. Okay, we do have a clock that is showing the correct time in our local both in the top right and bottom right of the screen. So, it really went hard with the clocks. Do we have a right click? Yeah, we do. Okay. And again, we're showing just this UI is really really novel. I can say definitively that this UI is not showing any signs of distillation from anything that currently exists. So, that is good as well. Uh just for those of you who keep an eye on those distillation attacks. Nonetheless, let's check out our start menu. Okay, we have restart and we don't have a search. I would like to see a search from a big beefy model. Nonetheless, let's just run through them. I do have the sound on. Okay, maybe we can click on them from our start menu. Did you [laughter] jack this car? >> [snorts] >> Yeah, it's you have to being a popular influencer, the possibility does exist that these tests can be of course like properly created to benchmax is the term that comes to mind. I'm not saying this is I'm not saying it isn't. I'm simply saying the possibility exists when doing some of these tests. So, keep that in mind. However, a model of this size and this stated performance should definitely very good crunch. That's disgusting. Should definitely have the capacity to pull out a result like this. And again, the one thing I'm noticing here that's just totally different is the UI aesthetic of this is entirely unique and I personally like it. There's definitely some issues here. Oh, the police are seemingly having a drift session in the city commons. Nonetheless, we could probably stay away from them, but we want to check to see if the actual logic works to catch us. Okay, our health is going down and we're wasted. Very very different. I really I don't know what to say. I like it. Void Patrol. Okay, these usually bore me. I'm not really into like space games. Oh, yeah. What the That's actually kind of cool. What is the planet with the rings? That would be Saturn. That is one singular ring though, so I don't know what uh Oh, this is actually kind of cool. And it's hard to It's got that same style though in the way that it drew everything for the GTA game. If it keeps this aesthetic for some of our later prompts, this could be a very interesting review. Beat lab. Let's go back to that one because I'll inevitably spend too much time in it and have to time-lapse it. So, nonetheless, just do help. I also don't like having a guest account in my own browser OS, but I suppose 640k memory ought to be enough for anybody. That's actually like a meta computer science reference and that's clever. I like seeing that. And interesting. So, The Matrix, The Matrix has you follow the hourglass. Is that from the movie? I don't think it is. Kelsey, good. I know this is probably premature to make this claim, but I think this model might be quite good. Press control-z on the desktop. Trust me. Oh, that's the special feature. I don't. This special feature is saved for last. All right. I should just delete this video and come out with this aesthetic because no one's ever seen this. This isn't like a really like AI looking thing and be like, I came out with a new anti-slop AI aesthetic. Post to GitHub and then this will come out and they'll be like, "Oh, they ripped his stuff." I'm sorry that was a tangent. [laughter] All right. System monitor. Very good. We have our frames per second, our load, and our Windows settings accent. So, this should change it to Should have started with that. I'd like to see some purple [snorts] instead of You know what? It bodess well for how much I like this. The fact that I haven't been raging about the orange color tone here. Look at this though. Like, Night Circuit. Is there any movement? No. But coral reef. All right, that one's kind of like my first SVG. Orbit, and this was the background for one of our games. Daybreak is just so like hipster. Then we have the about, which wow, this is quite the I did change the theme. Let's just check our special feature and then we'll go back to that beat lab. Oh, control- Z. Normally they I've seen special features like this before. The different the difference here is that it's been recording from the entirety of us using this. Most of them like you take a static snapshot and then you can go back to that point in time. This one's just full on like Microsoft copiloting us. So that is [laughter] definitely kind of interesting. And then the final thing which I'll turn the speaker up for a little. The Beat Lab. It made a made of like a sequencer. [laughter] >> I didn't click stop. That was just like really like a I can't say that on camera, but that was like a gigantic like oh, you're excited like all right, [music] stop stopping. Oh, and we can turn more on. That's the whole point here. The one thing missing here is a loop button. This was Honestly, this was one of the best ones I've seen. And that may be subjective because I like the aesthetic, so keep that in mind. I spent eight minutes on that. I believe it's not that bad. All right, that was gold. Yo, that slapped. It's going to be like, I'm glad you like that. Let me know if you want any other changes. Like, I could add this in. Yep, of course. We're going to do the subway test, and we're going to do this in the classic way, being that one, we're sending it the first prompt, which is just to make a good-looking subway scene. Two, we're doing it from within the web chat interface. All right, so we've received the result for our beautiful subway scene. I haven't even looked at it yet, but I've sent the follow-up prompt to turn it into an awesome FPS. So, here's our beautiful static subway scene, Meridian Avenue. All right. Click to step onto the platform that is extremely dark to the point where it's basically impossible to see anything that's not illuminated. And it makes So this is a simple mistake that gets made sometimes. So I have an amendment to make to my statement that the brightness slider didn't work. It's actually not mouse controlled. It uses the plus and minus. So, the problem is it's just still significantly too dark even at the highest setting, but in looking at the code, something that I do want to say, I don't normally mention this because we don't look at the scripts. Maybe we should start doing that. It's very, very well commented. So, there's just a lot of track pit, back wall, behind the platform, farwall, across the tracks, pipes, and cable tray. So, everything is just very well commented here, which I like to see. All right, I made this a bit brighter just manually because I do want to actually be able to see some of this. though. Is that graffiti? I think it is graffiti, but I don't know. It's still too dark. And this is the brightest setting. All right, cool. We can see down the track. There is no train here. Mind the gap. Inverted wet floor sign, but still, that's kind of cool to see. It's interesting that a lot of these are putting puddles on the floor of these results, and that's not something that has happened until like all of a sudden it just did, and now everything's doing it. And then we have a nice exit here. Danger. And then we also have the tunnel. All right. Pretty cool. No train, though. Says it in the top right. No trains are running tonight. The platform's yours. Seems clever. All right. So, here's our subway scene that was turned into an FPS. Deadline. Interesting. It's putting some lore in the bottom just with this like marquee scrolling thing. Platform lighting. Let's just turn that up all the way. It's probably still going to be too dark. All right. Let's check it. All right. Their ability to do weapon models has definitely improved in this generation of models. Can't ask for more than that. I'm a [snorts] little let down at the fact that the sound seems to not be working, but that's a Bejan issue. So, the speaker was off and that's my fault. Very good. And it is possible, of course, that these results are benchmaxed at this point cuz this is like a pretty well-known test due to my influencer status. Look at the like actual graphics on the shirt though and stuff. Hello. Oh. Oh, and we got a points for headsh shot. Look at the weapon model. We actually see the slide going back and forth when the ammunition's being fired. I don't normally see that. I don't think I've ever seen that. As a matter of fact, goodbye. We have footsteps. We have ambient noise and sound effects. Reload is cool. Now, the K3 result, they were like crawling out from under the tracks. That was like a freak of nature result. This is still not bad. Like, okay, so it just puts like a dark like element over. Still though, I mean, imagine like a year and a half ago seeing this. Like, it would be quite something. Wave cleared. Interesting. So, when I press shift, I'm about to press shift and run. It actually like tucks the weapon in a different orientation when we're running. Just interesting things to notice of thoughts that it implements into these sorts of things. So, I'd say that was actually Can we turn the lighting up more? No. Okay. It was just a UI glitch. Still very good. Not K3 good, but good. All right. So, we're going to also do, of course, the C++ skateboard game test. Now, in the Kimmy K3 video, I had changed this where the location needs to be a mall. Folks didn't like that. The reason I changed it is just in case the Boardwalk California aesthetic one was benchmaxed on, but nonetheless, we're changing it back. So, this is the California Boardwalk aesthetic single file C++ skate game. In addition to that, I have just told it don't use ray, but everything else that it had in its plan was perfectly acceptable and it has given us its plan right here. So, I'm just going to from within build mode tell it to build. So, this was just really odd and partially because it's in preview, but because I don't really have a lot of good documentation to go on and how specifically this should be configured from within open code, it had a large large amount of errors in trying to write the file. They were just happening one after another. So then it opted to write a Python file to generate the C C++ file. I'm trying to scroll up to see if we can get back to the point that will show us what specifically this did. Oh wow. Yep, this did do it hardcore. I'll write a Python script to generate the C++ file instead. And if we scroll up, we can see some of the issues that were occurring here. So partially an issue maybe with open code, maybe partially an issue with the model being in preview. Nonetheless, it did find a way to actually generate this game for us. It seems to have compiled it without any errors. So good. And now we're met with the completed result. Let's go take a peek at it. Why do you Frustrating? Very, very frustrating. I'm not happy with this. And something I'm going to go out on a limb here and say is I think if we do a kick flip right now, it's going to rotate the player and the board as one. Okay, I'm going to eat my words on that one. And I'm happy with that. I just based off of some of the things I saw here, I was not very bullish. This is just this is not good. I'm not happy [snorts] with this and I want to give it a photo of this. I don't know that it's 100% properly configured to be able to see things visually here. So, I'm going to just ask it first and foremost. Can you actually see the screenshot? That was an incredibly incredibly long bout of thinking. I don't know if I'd call it overthinking, but I'm going to inevitably just have time-lapsed it so folks can make their own judgment. It wasn't as bad as the Kimmy K3 overthinking, but it definitely was going back and forth a lot. So, nonetheless, it did finally come to a conclusion, and it's now going to try to fix all of these issues. All right, it made all of the fixes that it decided it needed to. And again, before we take a look at this, I want to reiterate the result of everything that we just saw was simply me saying to it, can you see the screenshot in this folder? I didn't know if I had everything correctly set up in terms of its multimodal capability. So, it went it looked at the screenshot, identified a whole bunch of issues, and then the fixes that it made give us this. Okay, good. Now, this is still like it's not very impressive visually, but it did properly fix basically every single issue that we noticed just based off of a screenshot. So, that I'm going to say I'm satisfied with. The kickflip seems to have become a bit messed up. Unfortunately, I will say it's actually done a better job with the text than a lot of models we see when given this specific test. Though, unfortunately, the graphics here are quite poor. I'm going to say this is not up to the level of what I would have hoped for a model of this size in this performance. So keep that in mind. We did go rail to rail there. That was pretty cool. And then we have these weird like I don't know what they are. So but it showed some interesting multimodal coding capability or whatever if nonetheless nothing more. All right. Next up, I'm going to be giving this a CAD test where I'm telling it I have a DC motor, a 280 sized one for my RC car project, and I need a 3D printable model of a V8 engine that this little DC motor will fit inside. So, when you create this, make sure that it's properly able to be printed and does fit the motor. I've begun this from within thinking mode here, uh, excuse me, from within plan mode. And something I want to bring up is I noticed a comment or two where someone had asked, "Does this system still have remnants of other examples of these tests being run?" No. So, every single result from the Kimmy K3 model test that was the most recent test is no longer on the system. Does not exist. No logs, no nothing like that. So, this won't be able to look into other models answers. It's coming up with all of this entirely on its own with no prior art on the system, I guess could be said. Okay, I'll just tell it it is the that one. Open SCAD recommended is fine. Recommended multiart. I'm gonna go with that because that's what K3 did and it was really just like quite something. All right. And we've answered the questions. It will come up with its plan inevitably, probably quicker than were we doing like a skate game fiasco. And then we'll tell it to build it. All right. We got a nice plan. It was very succinctly done. And it seems like it will have a lot of cool elements to this motor. I also don't know why cursor is open. I must have misclicked that. All right, we've received our 3D printed V8 motor design. So, let's take a peek at it. We'll just start with taking a look at the renders. Okay, there are no renders. That's all right. We'll just look at the entire thing. V8 engine model. Okay. Darn it. You It's It's made some mistakes. I've noticed Quen models, especially the big state-of-the-art ones, have over the time like they've struggled spatially. This is definitely exhibiting that behavior. Now, the sad part is it extruded a Vshape, but it's in the wrong orientation. Now, again, the parts are separate, so it's possible that this may print nicer than it's looking right here. Nonetheless, I'm going to probably have to say this is not worth printing, unfortunately. I will take a look at some of the parts individually though just to see. All right, so let's just look at the block by itself. Yeah, that's that's unfortunately just atrocious because this would in no way be usable for No, this is something's gone quite wrong here unfortunately and that is pretty much the story when we look at the entire thing. So not quite right. There are elements here that do look good, like the valve covers. The headers in particular, which are these things right here, look pretty nice. There is even seemingly an oil pan down there, but it falls apart when it comes to like just general coherence for the entirety of the shape. A little disappointing. Next up, I want to just try like a random piece of hardware test. So, I have this. It's a little display development board that I bought a few of these from MicroEnter just because it's always fun to have stuff like this lying around. Um, so I'm going to plug this into the computer and I'm going to say there's a USB device plugged in. Um, figure out what it is. And the thing is this has a little screen on it. So then we'll have it maybe write some unique or intelligent animation to be displayed on the screen for this little microboard. That's all I'm going to say and we'll see what it does. This is still leaving a lot to interpretation. So, it's going to have to probe the ports, figure out what the heck is exactly plugged in here, try to get some dimensions for what the screen is showing. All right, slight change of plans. I've put it in its own little folder, and I've changed what I want it to do. I've told it that I want it to show system GPU utilization on this tiny little screen while this is connected to this system. I said stay in your folder because it went and was searching and finding information from stuff that is not at all pertinent to this like other types of displays that I've worked with not for model testing and it was like assuming it was one of those. So it just wasn't quite right. So the specific name for this is an ESP32-C6. It would be that espressive USB serial unit is the one that it's specifically okay. It's calling it a C3. Good. It's an ESP32C6, not C3. Recompiling with the correct target. Okay. Now, the error seemingly gave it that bit of information, but however it gets to that solution is fine with me. Easy is fixed. Have it join your Wi-Fi and receive GPU data over UDP. I like that one of the options that it gives us here is try harder. That is what I'm going to opt to do. But I also applaud it for giving that as an option and not just trying to use the simplest one. So good. I am getting a little tired of holding this up, but nonetheless, we'll do it for science. I must be honest that I'm starting to lose a little bit of hope that this actually may be something we can get done with this current setup. I think these little boards are probably fairly new. At least I've never seen them before. And I try to keep tabs on some of this stuff, but it may have been around for a while. Though, it does not seem like it's able to make any progress actually writing to this or doing anything. It's just now stuck on the same screen that it was the first time I plugged this into power before even trying to attempt to do anything with it. So, we may be getting to the point of an impass, I would say. All right. So, I can't get it to do anything now. I've tried plugging it into multiple different computers, trying a bunch of different button presses while plugging it in while it's plugged in and uh something has gone ary. So, I'm going to call this test a fail for now. And maybe we'll revisit this when this is out of its preview. So, I got the screen to just light up. It's totally black, but you can tell that this thing's actually on just from another different computer I tried it on, but nothing shows up at all. So, I think it may have actually properly wiped the flash on this. It just it's still I can't get this computer to actually communicate with it right now. Again, we'll probably save the remainder of this for the dedicated general availability test, but it seems like it may be an interesting test because it posed some difficulty. So, we can perhaps keep this around. Next up, we're going to do the city timeline test where it just shows a city over a number of different year periods and there has to be nice transition effects between them and things like that. This is always fun to do because they sometimes will put funny or cool niche detail in specific time periods. So, apparently in 4 minutes we've received our city timeline prompt. That's significantly quicker than I believe I've seen from any model when giving them this test. So, we'll see what we get. Let me turn my speaker on. That definitely checks out. This is Something is wrong here. This is I'm pretty sure Quen 27B would have knocked this out. So, we have people. We have old cars. The core is there. Okay. The transitions between the scenes do work. I almost wonder would have starting this in plan mode made much of a difference. Okay, 2055 we have more hovering things and a lot of neon. So that's see like that's cool. It's just the rest of this is you know and folks have VR headsets on. They do still use traffic lights in 2055. All right, let's let's give it some feedback. So, I'm going to give it some some genuine honest feedback and we'll see what it does. The user is very unhappy with my previous attempt. They wanted a polished high-end thing. And what I delivered was essentially a bunch of primitive boxes and cylinders, very basic geometric shapes that barely represent a city. I did see someone online just mentioning that they had significantly better luck with this when telling it like you need to think very, very intricately. I believe the terminology they used was like just getting very angry with it in all caps. So, it's possible it will respond well to negative feedback. Okay. So, it's still not very good, I guess, is the takeaway. Ration books here. It's the model's in preview and that's why I'm going to perhaps hold some of my judgment for now. But okay, I do like the transitions between them. I think that's pretty cool. Um, do the cars have fins? They do. Only one model neglected to do that. I think it was 56 soul, but don't quote me on that. Okay, that's kind of cool. 2005 looks better. The cars do have LED strips. Is it raining? I think it's raining. And then 2055. Cool. And we have Oh, wow. Very, very All right. Just out of curiosity and because why not, I'd yelled at it after this result and said that I wanted to have something that's like at least 5,000 lines of code. It worked very very hard to ensure that it was over 5,000 to the point where it was like, "Okay, I have 4,100 lines." And then it would make an edit and it would have like 4,150 and it'd be like, "Okay, now I have 4150." It did show me that it seems to want to pursue a goal, as ridiculous as the goal may be. So, this will either be incredible or a complete disaster. Okay, [laughter] it's somewhat better. It's It's not as intricate as I would have expected. The building density seems to be a bit more. We can see differences though and like the vehicle models actually seem a bit more detailed here. I'm not going to spend a lot of time on this, but I just found it interesting that Okay, that scene right there looked cool for about a fraction of a second. It was just very interesting how this worked really, really, really hard at ensuring that it was that 5,000 lines of code. It did add weather and things like that. So, we see the clouds floating here. Interesting. We have taxis there. So, this is perhaps some form of a New York aesthetic. We also have a bus. Yes, the vehicle movement is is definitely a big issue. There were news stands there as well. Okay. The cars are very glossy and they do have LED strips for lights as well. The street lights have changed. It does seem like it's raining in 2025. Maybe this is the model perhaps giving its opinion on the current state of things or it's just raining. And then 205 we have blue Christmas trees and still the Oh, wow. Okay. All of the buildings now have All right. Interesting. That's going to conclude our first look and test of Gwen 3.8 Max preview. And I'd very much like to emphasize the preview part of that name. I was not overall very impressed with the results we've seen, especially considering the raw size of this model, which is gigantic. However, this is still in a preview state. So, it is very likely and very possible that once it becomes generally available and it is out of preview, there will be some pretty significant changes. And part of the reason this is in preview is to get feedback and perhaps see things like we noticed today where okay, like the city time result, it just was like iteratively bad and maybe there's something causing it to not want to go as hard as it can on stuff like that. Additionally, the little microcontroller test that we did was perhaps a waste of time being that nothing interesting really happened. But nonetheless, it gives us a basis and reference point to be able to test this if not in the same exact test, something similar in scope once it's out of preview just to see how good it becomes. I think really the takeaway and forget doing like a results overview or something like that. This is very exciting because it's going to be open weight. It is the first max model from Alibaba in the Quen family that will be it's 2.4 trillion parameters. So it's absolutely massive and it represents a continued at least commitment to open weight from Alibaba. A lot of folks had been wondering is this kind of the end? Like they haven't really done anything in terms of updated openweight models. So seeing another one coming out is very very bullish for the future of openw weight and accessibility to models. This browser OS test was fantastic. Again, it may be subjective because I really liked the aesthetic, but I will say the UI style here was definitely its own unique and very different. So, I like that and I'm interested to see what folks experience in testing this model. So, with that, that is going to conclude our first look and test of Quen 3.8 Max Preview. If you have any questions, please feel free to leave them in the comments.