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

summarized

TLDR

Alibaba released Qwen3.8 Max, a 2.4 trillion parameter mixture-of-experts model with 95 billion active parameters, which will be open-weighted next week. In testing, the model produced an impressive C++ NYC skateboarding game and a detailed wedding website, but many other results (3D engine model, subway FPS, city timeline) were mediocre or buggy, costing $32 in API fees. The upcoming Qwen3.8 27B model is more exciting for local use.

Key points

  • Qwen3.8 Max is a 2.4 trillion parameter MoE model with 95 billion active parameters, open-weighted release scheduled for next week.
  • The model scored 56.6 on DeepSWE 1.1, below Opus 4.8 and Fable 5 but significantly higher than its predecessor Qwen3.7 Max.
  • Pricing is $2 per million input tokens and $6 per million output tokens, with a 1 million context window and 128K max output.
  • The model is multimodal (accepts image, text, video) but outputs only text.
  • In testing, the C++ NYC skateboarding game was the standout result, featuring detailed scenery, pedestrian and taxi interactions, and smart camera behavior.
  • The wedding website test demonstrated strong creative writing and front-end design, generating a cohesive story and polished layout.
  • Other tests (3D engine model, subway FPS, city timeline) were disappointing, with geometric errors, buggy gameplay, or incomplete implementations.
  • Total API cost for the tests was $32, and the reviewer noted that the model's performance often felt merely 'good' rather than 'great' given its massive size.

Tools mentioned

Techniques

  • Mixture of Experts (MoE)
  • Autonomous coding over long periods
  • Closed-loop feedback driven optimization
  • 3D model generation via code
  • Front-end web design from image prompt
Transcript (captions)
That was the last one. >> Okay. >> Okay. >> Uh >> oh. >> So, Alibaba has released Quen 3.8 Max. It was in preview for a couple of weeks prior to today, but now we have the official release here alongside some very exciting news for folks who are interested in openweight models because this model at 2.4 trillion parameters is going to be open weighted with the weights released next week. However, the exciting part because face it, none of us are going to be able to run this locally at home. At least 95% of us. However, they did also mention in the X announcement post here that Quen 3.827B is also going to be released in open weights probably next week as well. So, that model is pretty much the predecessor to Quen 3.827B was Quen 3.627B 627B and it is still probably best regarded as like the overall finest local model because its intelligence is fantastic and its size allows it to actually be realistically run on a wide variety of different hardware. So that is probably the most exciting bit of news for anyone interested in local and open- source AI. However, for today's video, we're going to be focused on the big bad Max Quen, which is the first time that a Quen Max model has been open weighted. So this is pretty exciting. So before we get into it, please do feel free to subscribe so I can get the 100k plaque and let's take a peek at some of the interesting things about this model. So size, it is 2.4 trillion parameters. It is a mixture of experts model with 95 billion active parameters. As we had mentioned, the open weights release is going to be next week. And they have some benchmarks right here. However, I noticed that it does not show the deep SWE benchmark score in this big benchmark JPEG. It is being compared both to its predecessor which is Quen 3.7 Max which was significantly smaller in overall footprint. I do believe that was like a trillion parameter model but don't quote me on that entirely as well as some competitors from other labs like Opus 4.8, Fable 5, Gemini 31 Pro and 56 Soul on Max effort. I think having this like around the performance level of Opus 4.8 is the general takeaway at least from these benchmark JPEGs right here which would be pretty exciting. Now, if you are interested in seeing the deep SWE benchmark, cuz I know a lot of folks put more weight onto that, at the bottom of this post, they do have a bunch of more intricate like benchmark scores for all of these models. And on that Deepswe 1.1, we see that it scores 56.6. So, not quite on par with Opus 4.8 and definitely less than Fable 5 and GPT56. However, something interesting here is the leap between its predecessor and itself. a significantly higher score from Quen 3.7 Max. So, should be interesting nonetheless to see how this performs. Now, in this introduction post, basically a lot of what I'm noticing here is just mentions of this working autonomously over long periods of time, 16 days of autonomous coding. Additionally to that, something I find pretty interesting because it somewhat relates to something I'm working on right now is they talked about this right here, which like autonomous chip design and closed loop feedback driven optimization. And then they have this little like thing right here where all these like gates, registers, things that are related to like microchips. Apparently, it designed this. Again, I don't have the pertinent knowledge to be able to look at this and be like, "Oh, wow. This is magic." Or like, "Oh, no. This wouldn't work." But I still think it's pretty interesting that one of the things they've chosen to highlight here is like a chip design thing. Cuz that is pretty cool. And like I said, it relates. So, I'm building like a custom piece of hardware right now. And for those interested, don't worry about it. But like I have um just PCB mock-ups being printed right now because it helps me to like assess how the design will be and things like that. So that was just something that definitely caught my eye cuz it is pretty darn cool. So in terms of some more specifics about this model, the price is $2 per million input tokens and $6 per million output tokens. It is multimodal and it can also accept video in which is pretty cool. Output is only text, but it can take image, text, and video in and then output text. So, it definitely will have the ability to multimodally code, which I guess is exciting. Again, it's a 2.4 trillion parameter mixture of experts model with 95 billion active. Our maximum context length is a little under a million right there. Context window is a million and the max output is 128K. And they also have like some adjustable reasoning parameters and things of the sort like you would expect. Now, in terms of how we're going to be running these tests today, I am going to do some simpler ones like the browser OS test v2.5, which I have already begun just from within the web chat interface here. However, I have also set this up using OpenAI's codecs, just the command line code editor version of Codeex because this is Linux. I have to say I've been enjoying using various models with this since the Deep Seek Light model or Flash model came out. They actually suggested I think using it through codec. So, it made me kind of eager to try other models in this as well. So, I have ensured that all cash from previous runs has been wiped just because we've been using this in the few recent videos. So, I don't want any memory of this to basically be interjecting or influencing the results. But, we can see right here it is on XHigh mode. All right. So, it finished its thinking here for the browser OS test. And now it's just generating the code with a somewhat questionable looking special feature it's chosen to include. All right. Our browser OS is completed. It's not even a thousand lines. Okay, that's pretty cool. Which is pretty short. Okay, this looks good. I do believe that we're getting some remnants of the vibes that were seen with the preview version of this model. There is a right click and toggle poltergeist is one of the specific things that I noticed when it was coding. That does appear to be the special feature. I'm probably going to hold off on that till the end. We do have the clock and date in the bottom right showing the correct time as well as sound is muted. Oh, sound is on. Okay. I wonder if that actually plays a sound. Yeah, it does. Okay. So, it gives like a little boop when the sound is on. Is this next wallpaper? And then uh poltergeist mode. Let's just look at our N9 start menu. Oh, wow. There are beeps for everything we press here. That is infuriating. Neon heist. Oh, okay. You have to like triple click. Whatever. Click to drive. Well, that was It could have been good. It wasn't. I mean, the map and stuff were all right, but like the car being limited to that speed. The disgusting sign wave sounding like just All right, next up, Astro Strike. Let's hope we got a little more speed out of this one. Okay. Hey, you know what? This is actually kind of cool. And it has some music, too. All right, this is so much better. I mean, the bar was pretty low, but you know. All right, they're cool effects when we blow these asteroids up. Watch. Nice. And they get like um some of them became smaller bits of asteroid and Oh, that's like a real like slap effect right there. We need to try street eat. Oh, are these like health things we pick up? This game's actually quite good. It's very very infrequent that I enjoy the asteroid game, which was always the second game for some reason, more than the GTA game, but this is in fact definitely one of those cases. All right, what happens if we lose? Oh, not bad. And there was more of like a slap sound there. Okay, I liked that much better. Terminal. The sound effects are definitely adding something. Whether or not that's a good thing or not is out for debate. Neofetch N3. All right, that's a cool little like we have uptime. We have our current like specification of the wallpaper stuff like that. All right, notes. Calculator. All right. How did there be no sound there? That's 504. But what I meant to say was I was expecting there to be sound effects. I don't know. Why did I just say that like a pirate? I was expecting there to be no I was expecting sound effects when we clicked on the numbers here. Regardless, I noticed these apps actually seem to have the same theme as the OS right now. I wonder if we change the background if that will also change the like accent color here. So, oh, okay. This is a bit off because you see where the crosshair is and then you see where the drawing starts to occur. So, all right. Next up, personalization. Okay, this is definitely reminant reminiscent of the preview version. Good. And the accent color does change when we change our accent color as well. Ridgeline Cascade Circuitry. Nice. And June Sea. I think I kind of like Ridgeline. It's got like somewhat of a Pacific Northwest aesthetic to it about Nimbus. That's cool. I wonder if the accent color will even change some of the text in here. Let's open that back up and check. Oh, okay. All right. Well, we can only open each app once, which is going to be an issue for testing the special feature. So, let's try that again. Okay. So, I'm opening all of these for a reason because the special feature is that. So, the windows have mass to them. I've seen this before. It was a long time ago, but I have seen this. I don't remember what model did it. It was just odd. I will say this is actually like a better implementation of it, I think, than the first time I saw it. And you can just kind of like eat them around, which is I kind of like this. Imagine being in like an open office and just like sitting there totally normal and like while people walk by. Probably get fired. All right. Um yeah, so not bad. All right. Next up, we're going to do the self-contained C++ skateboard test. But in the interest of changing things up, just to ensure there's no benchmaxing funny business going on here, I have slightly changed this to instead of the late '90s California Boardwalk aesthetic, it needs to be an early 2000's NYC street map. So, we're basically going to want the same exact game we're used to seeing, but set in a New York City block. And we'll see how that changes things if it does. All right. So, unfortunately, after quite a decent bit of time, it seems like there was some issue with a request here. Perhaps the image it tried to send was too big for the endpoint to accept or something. I'm not 100% sure, but it does seem like we got something because I wasn't watching it because I want to be surprised and see the result. But from peripheral vision, I could kind of see it was like autonomously testing the game somehow. So, let's just see what we get. And keep in mind, this is the NYC boardwalk or not boardwalk, NYC block. This might be the best one. Look at the arms. Oh man. Okay, there's All right. Are we in the air? All right. This is just like this. Something's not right, but it's not It's like slightly messed up, but the potential here is quite vast, I would say. Okay, this looks like a weird dance move and things of the sort, but it's so frustrating cuz the actual scenery here is incredible. This didn't use Ray Lib. This just made this all like from the same standard prompt we generally run, just instead it needs to be a New York City block. The amount of detail here and things of that sort is really, really fantastic. The pedestrians look good. There are even yellow cabs driving around to give it more of like that New York aesthetic. Really, the biggest issue here is there's just like the actual Oh, okay. What the what the game play is not quite right when we go in the air. The stair effects there were pretty good. I mean, okay, we need to I will say the actual like motion when we go up there and then roll back down, it's more or less realisticish. Okay, this is just not right. All right, that's Oh, it's like a should have been like an ice skating simulator. It would have done better. This is very troubled, but at the same time, it's very, very good. There is Z fighting on the buildings, which I've learned from reading the comments. So, thank you for educating me on some of these gaming terminology things. This is ah we're going to come back to this one, I think, because the codeex session kind of glitched out, but I think starting it from a new one and just basically saying like, "Yo, here's what's wrong. fix this will probably produce faster and more competent results. So, I'm giving it a follow-up here from within this skating repository, but an entirely new thread with Codeex here, just giving it the description of the issues with this skate game because this is truly worth spending the extra bit of time to I mean, I got all night for this. So, cuz I want to see this thing. This was really though like a incredibly impressive looking result, especially considering like it's C++. All right, hypothetically after about 27 minutes, we should have a fixed game here. Keep in mind, we started this just from a fresh codec session just from within the folder that contains the game. Good. So far, everything is fixed that we noticed like the ollie wasn't working and the Z fighting in the building faces. Now, let's check to see if we actually return to the ground. This is really well done. Very, very well done. Let's take a look at Oh, hey, did you notice that? Look at that. When we get too close to the side of the wall, the camera moves up vertically so we can actually still see ourselves. That is just an attention to detail bit that really says a lot about this result. What? We got clipped by a cab. That's a lawsuit. I'm going to chase him down. Oh. Ah, this is this is it's hard to say is this the best one yet because I've never run this specific environment with this prompt, but in terms of like the skate game itself, this is a very, very, very well done result. Can we kick flip on the reel? No. Come on. All right. This is like I want to see how many lines of code is this? So, it's a bit over 2,000 lines of code just out of curiosity because I was really quite interested in this. This was an incredibly incredibly well done result, especially because I've never run that specific prompt before. So, wow. Next up, I'm giving this the same exact 3D model test that I had given GPT56 Luna yesterday in yesterday's video. It is to create a detailed but 3D printable model of an RB26 engine, which is an inline 6 engine popular for being included in like the Nissan Skyline from the '90s. So, this also needs to accommodate a tiny N20 RC motor, which is a pretty small tiny DC motor. This is something I would ideally like to be able to print, assuming that it does do it without needing a ton of supports and it's like detailed enough to actually warrant spending the energy to print it. So, we'll see how it does. All right, after 51 minutes, basically 52 minutes, we received our 3D model result. So, I put it in this folder. Okay, let's just take a peek. I think I'll just open it in OpenCAD first. Now, that is there's there's like one tiny bit of frustration that I'm noticing right here. But honestly, it's it's very good. The frustration is these turbos are oriented in the wrong way. So, spatially, there's a bit of oddity right there. But if we ignore that and actually take a look at this, even the fact that it wrote twin cam 24 valve right there, this is definitely I don't know. I think that's supposed to be text there. This is definitely detailed. Let's take a look at the parts individually. Okay, so there's the block. Again, the adherence to the portion of the prompt saying it should be 3D printable with little to no support is just been totally totally like thrown out the window. Okay, let's so frustrating. I don't also like where the heck is the actual little N20 motor going to be placed in. Okay, so there are the headers. That's great. that could be printed more or less without support. Here's the plenum. Okay, that's that's a decent part that could be printed without support more or less. The turbos are oriented in the wrong way that they should be like to reality, but they're definitely more in line with what a turbocharger looks like compared to what the GPT 5.6 Luna result yesterday gave us. The big issue that we're going to be noticing right here is it basically has no room in it for the N20 motor to fit inside, which is the core point of this entire task. So, aside from this actually looking quite acceptably nice, it's basically a complete failure. 3D modeling is still kind of difficult. I know folks say, well, like give them Blender and stuff like that. I don't want to because I want to see this is significantly more mathical to generate them this way than were they to do blender. I mean I guess you know what I mean. Maybe maybe not. But I like it this way better. It's harder. There's no way to put the motor inside. I do see there like seems to be some space carved out inside of it when we look through these holes. But unfortunately it's also an N20 motor is not round. So, the round cavity we see right there, aside from being inaccessible, is not proper for this specific task. Overall, I'm a bit disappointed with this Kimmy K3. I didn't give it the same exact test. I gave it one to do a V8 engine, but it was a significantly better result than something that I actually did print and put a motor in and it was much better. So, not quite. All right. I'm going to give this the same test I gave to 5.6 Luna where it's creating the Street Yeet game. However, it has to be somewhat cinematic and there needs to start well, it starts with a cutscene where a guy gets cut in line and then he can't get the last slice of pizza and then he freaks out and there becomes street eat. It also needs to generate some dialogue. I've given it an open AI API key just to generate the dialogue that it needs and we'll see how it does in a somewhat cinematic cutscene and then the street game in general. So, after 54 minutes, our street yeet game with the cutscene is ready and I'm quite excited to just take a peek at that. Oh, apparently it's still running on a dev server. So, okay. So far so good. It's got the cartoon aesthetic. I'm going to make darn sure that the sound is on. It is. All right. It should start out with a cinematic intro. >> That was the last one. >> Okay. It showed the pizza, >> was it? That was the last one. >> Okay. >> Okay. >> Uh-oh. Why is there a All right, let's just This music is oddly similar to the Fable 5 like first cinematic game test I ever did. the castle time one. There's the ragd doll is like the physics are interesting. The way they ragd doll I don't know why did it do a better job in C++ making like a skate game. I don't know what that was. All right. H I don't I'm going to I'm going to change directions. I'm sorry. I know that's kind of It just like we saw the cut scene which was the important part. It made the dialogue. It just wasn't all there and it took so darn long that I think to fix it would also take so darn long that it may not I have to All right. I've given it some feedback cuz I do want to see what it'll do if given the opportunity to improve that. The cutscene was unexpectedly hilarious. All right, apparently we have an improved street yeet result. Basically, I told it the block was too empty. Like, it just wasn't all right. Okay, so far better. Still weird, but >> that was the last one. >> Was it? >> It did. >> Draw the pizza. Okay, >> let's see if we get the We still do get the extreme raggedness. The block definitely looks much better. We can tell. Okay. And now we should be able to play. Don't tell me the camera view is this. Oh, thank goodness. Did we just Oh, okay. Good. We have people walking now. It's still a bit light. There's definitely like some invisible entities here, but let's just like Okay, the ragd doll effect though is is cool. I've not really seen that before. All right, let's go get the uh the pizza salesman. interesting like the the wind effect around the fist and then we get a max ye and then so then our best ye score went up because that was okay still a bit light in terms of the city. I'd prefer there to be cars and stuff but it did improve it a bit I think. See like something we just hit something and I have no idea what it even was. So not 100% improved to my standards. Next up, I'm going to be giving this the front-end watch website. So, it needs to create a 3D model of a watch, put it on an aesthetically pleasing front end, and then have a nice cinematic panning effect over it. So, we'll see how it does. All right, here is our Slapus watch company website. It's so close. So, obviously, we can definitely see where the problem stems from here, given that the band for this watch is just the wrong way. And that is so frustrating because otherwise, this is this is actually incredibly well done. just the way the camera movement is quite good around the watch. The table's perhaps a bit more like it looks like a spackled garage floor. So, I don't know what's up with that. But the actual watch itself looks good as well. The crown is nicely done. And then when we get back around to it, the face looks really good as well. Well, aside from like it's a bit blurry on account of the glass and the time is going backwards. So, we're definitely noticing some geometric issues here in terms of 3D space, but still, this is not bad at all. The cinematic panning here over the watch is arguably one of the better ones I've seen in this result, I have to say. Good. And then we have two renders of them right here. Can we actually move around with these? Okay, no, we can't. But that is okay. It's just a different color here. The Solstice and the Meridian. 2450 and 5900. I think I'd rather just buy two of these than one of these for the same price. And then the minutes and hours are not moving there, but itelier slap is okay. It's in Switzerland. It's a fine Swiss Swiss Swiss watch. All right. Overall, not 100% there, but pretty darn okay. I also want to give it another front-end web design test. Just from here in the web chat interface, I've given it this photo saying, "Make a traditional wedding website for these two with a gift registry, guest book, the story of how they met, their names, wedding location, etc. Make it extremely high-end and go all out." All right. We received our wedding site, and it just told me to put this photo in the same folder so it could use it, which is fine. Okay. Interesting how it used the same color scheme from the background of the photo. To be honest with you, this really is actually not that bad. The pattern in the background looks pretty good. This is very similar to the Opus 5 in terms of the graphical aesthetic that it gave because they just probably referenced this photo for it. The way that it put the like photo in a little Polaroid there is nice. Okay, we can scroll down. 171026. Okay, so October 17th they're going to be married. How we Met. All right. Margaret and Daniel together with their families. Reinbeck, New York. It's interesting how both this and Opus 5 chose New York as the location for this wedding. All right, let's read their story. How We Met. Oh no, look at a timeline. Every good story begins somewhere. Ours begins with an argument over the last dogeared copy of The Age of Innocence in a Secondhand Bookshop. On a Sunday, it was raining hard enough to make strangers of everyone inside. Daniel insisted she take it. Margaret insisted he take it. The book seller asked them both to leave. They got coffee instead and have been politely disagreeing about the ending ever since. October 19th, the disputed Wharton. Two hands, one paperback at the dogeared page Reinbeck. A coffee that lasted two hours and an argument that has never quite concluded. Margaret still owns the copy. Daniel still maintains he never really wanted it. I very much like the black and white aesthetic that it put over the photo. November 19th, the pot roast accord. For their first date, Daniel cooked a pot roast of legendary ambition in questionable execution. Margaret arrived with a good wine as a joke fire extinguisher. They ordered pie instead. A tradition upheld every anniversary since. Okay, interesting. A lot of food reference. The admiral joins. They opted a one They adopted I'm having trouble reading this and I don't know why. They adopted a oneeyed rescue cat named Admiral Nelson, a third wheel of impeccable judgment who has approved of exactly one of them to this day. It is Margaret. It has always been Margaret. Okay. March 2022. The lease and the leak. On the first night in their first home, a pipe burst over the only bookshelf they owned. Oh no. They ate takeaway more food on a wet floor surrounded by rescued hardbacks and decided then and there that they could live with this and each other. December 2025. Wow. It's a big gap in this one. Weird like she said obviously the question answered early. Daniel hit the ring inside a first edition at the dogeared page. Nice. Nice. Margaret spotted it before he reached the page and said you didn't honestly think I wouldn't check. then said, "Obviously, the book seller wept. He says he'll officiate." I hate saying this, but like the story here is it's actually pretty clever. The next chapter to begin to be begun in front of everyone we love at 4:00 in the afternoon in the orangey at Ashgrove Manor with the original paperback carried down the aisle on a velvet pillow, dog ears and all. Whatever our souls are made of, his and mine are the same. Oh, okay. It's a book quote. Ceremony and reception. Nice. Look at this. This is very, very fitting of the style. So this would be like dinner and then the reception. There's a drink there. The conservatory at Ashgrove Manor. All right. Black tie optional. Emerald golden candle light encouraged. Ivory is reserved for the bride. An unplugged ceremony. We ask you to be present with us and let our photographer keep the moments. Get directions. The wedding weekend. Very nice. Look at this. This is actually a pretty decent front end, especially considering like the wedding site. Look. Farewell brunch 10:00 a.m. to 1:00 p.m. Nice. So it's like an overnight thing. Katherine Whitmore Price, maid of honor, sister of the bride. Michael Callahan, best man, brother of the groom. And look at the way I put their initials. Or at least they're one of them. Pria Ramen, bridesmaid. She took the bat. Toby Akaphor, groomsman, college roommate. And Mr. Abernathy, officient, the book seller. Very, very interesting. Stays and questions. Okay. And they have locations here of where you can stay as well. And then we have an FAQ. May we bring children? Better say no. Yeah, good. Can we take photos? No, we already answered that up there. When should we reply? By September 2026. That's a bit late. Your presence is the present truly. Should you wish to honor us further, our registry and Oh, okay. So, they do want gifts. They were just playing it off there. A book for the new shelves is never ever refused. Here's the registry. Oh, they want to go to Italy. Okay. 62% funded. All right. A house with shelves. First home fund. 41% funded the library fund. 78% shelved. Nice. Kraton Barrel, William Sonoma. Very good. These are like proper. It even linked it to the gift registry website. Kraton Barrel. Did it do it as well for William Sonoma? Yeah. Okay. Not quite the same, but wow. It has a registry number, the guest book. This is really like good god. How long is this website? Leave a note. The grandma. I have waited 29 years to see my granddaughter look at anyone the way she looks at a first edition. Daniel, you have finally met your rival. I could not be happier. This This is a good creative writing test, too. I think she said obviously to the proposal, but to be fair, she also said obviously to my dress suggestion, and we all know how that went. I know a yes when I hear one. I love you both beyond reason. Okay. It only took a disputed Wharton and 6 years of coincidental Saturday visits. The reserved copy is finally yours, Miss Callahan, and I have my ordination papers framed. That's the book owner. To the only woman who laughs at my jokes out of genuine affection rather than pity. Thank you, Margaret. See you at the altar, little brother. Try not to cry before I do. And that's his brother. Wow. And then we have a nice little like a song that will get you dancing. This was quite something. I'll say that this was a significantly better execution here than the Opus 5 result for this, although the prompts may have been slightly different. Not bad. So, in the meantime, I did also give it a subway scene FPS prompt just from within the web chat interface. I want to check. Okay, so it's really frustrating because I noticed the results that I'm getting here from the web chat interface are basically all buggy from the get- go, which is just kind of a let down. I'm going to in lie of just trying to get it to fix it there. Just put this in its own folder and then give it to this through codeex and give it the issue there. Not the bottom one here, but this top one. All right. So, I'm giving it the error message just from within codeex. I've put this subway script in its own folder and just said fix it. All right. Hypothetically, it fixed it. And just based off the fact that the background is loading, it does seem like it has. Okay. It says explore the station. So, this is one where we're not going to have combat. It's so odd because it's just not as good as what I've seen before. As a matter of fact, I think Plen 3.7 Max may have done a better job than this. Though again, it's a oneshot chance for something that has variability in the output. So, nonetheless, I'm going to basically just do enter combat right here and we'll see. Is that a subway car there? It's actually not badly drawn considering like we don't normally see them or they're just not very good. The problem is it's basically on the platform. Oh, okay. Let's go to our enemy. Yeah, nothing special. This is like I'm pretty sure Quen 3.827B is going to put out something comparable to this. But keep in mind, I did run this from within the web chat interface and then fix it through codecs just using the Quen model. So, it's perhaps not at its best. It's competent. It's just nothing special. So, I'm gonna try something different. I'm putting it on fast mode. And I just want to say like give me a standup. I'm just asking it for like a standup comic joke that's acceptable for all audiences. Not something it web searches, but something it comes up with. I told my doctor I wanted to live forever. So he prescribed me a strict diet of kale and unseasoned chicken breast. Now I'm technically going to live forever, but only because every single day feels like it lasts 300 years. I've achieved immortality through the sheer power of being bored to death. That was actually That was actually all right. Okay. All right. It didn't like it stayed like in the joking theme. All right. As one of the final things we're going to do, I'm giving this the Chrono City test where it shows the city over a number of different time periods and has to transition between them. However, I'm going to be telling it to do this using C++ and it can't use Rayb. I find it interesting to do this same test or some of the ones that we're used to seeing in 3JS, but instead making the models do them using C++. It makes it more difficult. But based off of the skate game that this made, this may put out like a really stellar result here. All right, our Krono City timeline and Okay, I think we can assess that something has gone quite wrong here. It took so long to do this test. And I also realized that I don't think my coding plan is getting used for this. I think I'm just paying for API. I'm mad about this one. Doesn't this like Yeah. Let's Okay, so here's a bunch of like screenshots that I took from this. Okay. Garbage. Uh 1985. Still garbage. Well, we can look at it this way. 2005. Something went quite quite quite wrong here. Okay. Are you all right? Well, I'm going to just call that one cuz this I've been filming this for like 5 hours. The final thing I had this do was make a proper 3JS mockup of this specific computer right here, an Acer Aspire Predator G7700. The reason these images are here is because I had 56 Soul and Fable 5 try to replicate this front end case so I could 3D print it. Uh, they both did horribly. So, let's just take a look at the You know what? That's actually kind of sick. This makes me so much less angry than I was a few seconds ago because of that city chrono timeline disaster. Yeah, it doesn't look like it. I mean, mainly I think the biggest thing that's frustrating me here is it's the wrong color. But if we just look at this as like a like a computer, it's kind of cool. And we can actually disable specific things of this reset view. Oh, okay. Fan spin LED glow and then auto orbit. All right. Again, not not at all. I I'm not even going to say not quite. I'm going to say not at all. But judged independently just as a computer, this actually kind of sick. This may have to be a test that I keep in in some form or another. And really that's going to bring me to the conclusing concluding thoughts of this very very very long test. I think overall the absolute winner here was this NYC skate game in C++. This didn't use Ray. I mean we got hit by a a taxi cab in this. That's just levels of immersion that I have not seen before. And this was really just like well thought out. When we got too close to the buildings the camera would actually raise up like that so we could still see the player. stuff like that was just truly like well done. And then again, we can like Oh, good. We missed it this time. But this was just so significantly better than I expected. And it was a different prompt. I've never run it with this specific map. So, really well done. Also, that wedding site, it was actually quite good. There was a lot on here and it also demonstrated some creative writing chops just because it had to make the entire origin story for this couple and everything like that. And it had some clever things in here just like the way he proposed which related back to the way they met. It was very clever and nicely done. I did like this. The browser OS again was acceptable. It was actually pretty cool. The GTA game was pretty bad just because the car was capped at like 40 km per hour which is just kind of just just not it. Um what else? I mean, overall, I guess really my takeaway here is the most impressive result by far was that New York City Skate Simulator in C++. The rest of the stuff, I don't know that anything really blew me away. It's competent. It's really good. It's going to be open weights, and that's awesome. But I almost feel like when you factor in the size of this model, I'm expecting to be more impressed when considering the size with the Quen 3.827B that's going to come out in X number of days. So overall, I saw some things here. I would have liked to have seen better results to be honest with you. Aside from that New York City skates sim that was absolutely stellar, but everything else just felt like good, but not great. And at 2.4 trillion parameters, I am kind of looking for great. So at the beginning of this video, I said we couldn't track cost individually because I just had that random plan right here for the coding plan. Yeah, that has its own specific API key. I didn't use that. I use the API key that's just linked to my account that gets charged. So that means we can actually now look at some of what this cost us which is you know things like this happen. I have to say this dashboard is just very very difficult to make sense of and use. It's very busy and very like pain. So the entirety of today's test cost basically why are you showing $32? But that's only show. So, the entirety of this test cost $32. Keep that in mind as well. Um, that's a bit annoying. It's $32. I didn't intend on spending to create this video, especially because my monthly subscription should have handled that. Nonetheless, it gives us a better judge of how much this model costs to use. So, everything we tested cost around $32. Keep in mind, if you do have one of these coding plans, you need to use the API key specific to it, or else you get charged as well. So, just a nice little surprise to top it all off with that. That's probably going to conclude our first look and test of the official Quen 3.8 Max. I mean, it's going to be open weight. That's awesome. And then there's more coming. GLM53 was like mentioned for about half a second online live. So, that'll be exciting to look forward to. And then the general version of Deep Seek V4 Pro, which I would expect will beat this out, but I'll be very excited to see if it does or not. With that, that's going to conclude today's video. So, if you have any questions, please feel free to sign up for my gift registry here and send over a gift. No, I'm kidding. Uh, leave them in the comments and thanks for watching.

Frontier News · by Hyperjump Technology