Qwen3.8 27B Is INSANE – This Is the BEST Local AI Model Yet!

summarized

TLDR

Qwen 3.8 27B is a powerful local AI model that can run on a single GPU, showing impressive results in coding, game development, and 3D modeling, though it struggled with a complex C++ skate game. It's open-source and a strong alternative to subscription-based AI.

Key points

  • Qwen 3.8 27B has a 262,000 context length, extensible to 1 million, and scored 42.2 on the SWE-bench, a huge leap from its predecessor's 13.3.
  • The model is Apache 2.0 licensed, making it one of the most open, and can handle images and videos, including hour-long videos.
  • In testing, it created a functional subway FPS game with sound and enemies, and built a 3D model of an RB26 engine using OpenSCAD.
  • It generated a high-end watch website with realistic renders and product differentiation, and a city timeline test with accurate period details.
  • It struggled with a self-contained C++ skate game, requiring human intervention to fix bugs, but overall performed exceptionally well for its size.
  • The model was tested at Q8 quantization on an RTX Pro 6000 GPU, and also via RunPod for parallel tests.
  • It showed strong agentic behavior, like adding a screenshot hook to debug its own code, and handled complex multi-step tasks.
  • The video concludes that Qwen 3.8 27B is the best local model when factoring in hardware accessibility and performance.

Tools mentioned

Techniques

  • Q8 quantization
  • extra high thinking mode
  • agentic fixing
  • multimodal input
  • procedural generation
  • 3D modeling with OpenSCAD
Transcript (captions)

0:00 I had it do a street eat. It didn't fully finish, but coming very soon. Why did it just go from zero minutes remaining to coming soon?

0:17 Where is it? All right, so for today's video, and you're probably going to be wondering, hey, how'd you get that little plushie? Well, because I'm an influencer. So, Quen 3.827B has been

0:26 released. In my opinion, this is probably the most long- aaited and exciting release for folks who are interested in local AI. Why? Because the predecessor was the strongest local

0:37 model of its size and had the widest variety. Basically, it can run on the widest amount of hardware without needing to spend thousands of dollars on compute. So, let's start by taking a

0:47 look at Quen 3.827b. I will probably go through this section a bit quicker than normal because I would imagine a lot of folks just want to see how this stacks up in coding tasks and other things,

0:57 myself included. In terms of some technical specs for it, it has a 262,000 context length, which they say is extensible up to a million. Now, if we scroll up here, they talk about hosting

1:07 this on this Quen cloud service where by default it has 1 million context enabled. So, just to mention that as well. Now, if we get to the benchmarks, this is where things become quite

1:17 interesting because I always tend to now just look at the deepsw SWE scores because I find that my experience with all of these models tends to be pretty decently reflected by the scores I've

1:27 seen for this benchmark in specific, of course, in coding tasks where this over its predecessor is a gigantic leap. And that says a lot because this predecessor right here was still like an absolute

1:39 powerhouse, especially for the size and openness of the model. So, if it really has increased from a 13.3 to a 42.2, this may be like some serious business. Now, it is also pretty heavily, I

1:51 suppose unscientifically, putting the Mac down on Quen 3.7 Plus as well. And if my memory serves me correctly, Quen 3.7 Plus was like a almost 400 billion parameter mixture of experts model. And

2:02 even look how close the predecessor to this was to that. So that just again lends itself to how good this model was and how exciting this model is because a lot of folks are expecting that. Again,

2:13 this is almost trading blows with Opus 4.6 on Max. We don't have the specific deep SWE score for that which I would tend to put more weight on in terms of raw coding performance. However, it does

2:25 remain the fact that this is essentially like a Frontier model from a year ago that can now be run on a single GPU, which is just absolutely madness and exciting for folks who don't want to be

2:35 tied to subscriptions to access intelligence. In that same vein, the license is Apache 2.0, which is one of the most free and open ones, probably not one of the most, probably the most,

2:44 which is exciting as well. In terms of its multimodality, this can not only accept images, but also videos. Now, they do even mention somewhere here like sending in an hour-ong video, which I am

2:55 definitely not going to be trying in this test because I don't know what that would do for VRAM requirements and use, but it's just insane to have that capability as well. From STEM diagrams

3:05 and documents to our scale videos, which is quite nuts. Now, in terms of the way we're going to be testing this specifically today, I have opted to use a Q8 quantization, which I've just

3:15 pulled the one from Unsloth right here. Reason being, while this will not fit on a single 24 gigabyte card, which I know then may alienate some viewers because the performance may be degraded when

3:26 you're using one of these Q4 quads, I want to see this at its more realistically highest potential performance. I'm not going to pull the BF-16 cuz it's just probably not

3:36 realistic. I don't think a lot of folks will use that, but I think seeing how it's doing at an 8bit quant is probably important just to get a more accurate feel for how this stacks up to things

3:46 like Opus 4.6. six and things of the sort. So, I want to mention that. Additionally, I am using this on an RTX Pro 6000 GPU, which has doubled in price since I purchased it a few months ago,

3:57 which is just kind of ridiculous, but um it's fast and it has plenty of VRAM to max out the context, so I guess that's a good thing. All right, so here's our browser OS. I've had the

4:08 absolute worst time so far in testing this model. It's been like three or four hours since it's been out and this is the first time we're seeing the browser OS. Now, a couple of things. One, this

4:18 is a new version of the browser OS dubbed the browser OS test v2.7. It is the same except it needs to also include beautiful procedurally generated backgrounds with movement and then also

4:30 an email client. These were things that were suggested to us just in a comment of the recent video. So, thank you for that. They were a good idea. Now, let's take a look. Something I noticed right

4:39 here is it included this thing. And if we refresh this browser OS right here, we're going to notice it's named it Pulse. This was the exact same thing that Kimmy K3 did in the browser OS

4:51 test. I believe it named it Pulse as well. And it genuinely looked exactly like this thing. And if you clicked it, it did also Bit is following you. 8 seconds of

5:03 companionship. I'm going to have to refrain from commenting on that in a more Okay, we have a clock in the bottom right and the date. Is there a right click? Yeah, there is. Good. New

5:15 wallpaper and it should be procedural with some movement. Okay, it is. I have to say they're kind of ugly, but nonetheless, shuffle icons. Oh, very good. That is a feature

5:28 I've not before seen and would be okay with not seeing again in the future though. All right, let's check our start menu. That's pretty good looking. Settings, a bunch of stuff. Let's just

5:40 get started. Now, having to include an email client with this, I think, is kind of difficult because it's not actually going to be able to knock out and put like an email client in here that's

5:50 functional. So, it probably just has spam at Wow, it's clever. Pizzaron 3000, your quantum pepperoni is ready. Hey, we got new mail. Our pet emailed us. Follow me. The snacks are calling. Try crashing

6:06 into a car near me for science. Can I send it? It's It's definitely not like related at all, but just looking at its eyes right there after sending that email is kind

6:26 of meta. All right, let's let's check our notes app. Probably going to be simple. New. Okay, word count, character count. I don't see the ability to save this as a text file, which is

6:38 enraging. All right, let's see what's up. Type help. I would if I could click into the search box or the text box. I'm going to be very sassy here. Pet. Ah, pet feed.

6:50 Oh, it's walking to the pulse storm. This video is not going to be out for like six years. Okay, this model is going to be like deprecated by the time

7:07 this video comes out. All right, I don't know exactly what just happened. Matrix. Oh, nice. Let's see if pulse has to get out of the way of these things or if it gets

7:19 no. Okay, good. Good. Nothing happens to it. All right, Neoch. Beautiful. Matrix just to turn that off. What else? Pulsecom. Okay. All right. We're all

7:34 curious. Let's check the GTA clone. Grand Theft We Oh, I can already tell this is sick. Look at that. I've not seen traffic this well done before. Wait a second, though.

7:47 Can we actually get out of the car? Also, Pulse is still here. Like, we're just going to ignore that. Oh my. This camera movement is is along the incorrect axis. And I don't

7:58 think we can get out of the car. This is in some ways worse than its predecessor, but in some ways better, this specific result, because the predecessor had like some

8:09 sick walking and stuff, unfortunately. pick up the package. Yeah, I would if the car moved. Oh, wow. We can maybe kind of like This is like a like a quantum space.

8:24 I feel like we just ran over a pedestrian cuz I heard something, but I I can't see because of the camera being totally totally messed up. Okay, let's let's come back to this or not. Next up,

8:35 we have Void Hunter, which we're all going to imagine is probably the same generic asteroid shooting game that all of these include. Actually, the movement here is it's pretty well done.

8:50 All right, this is cool. Oh, are we getting shot back at? No, I think we are. This is actually pretty well done. Can we move forward? No, but

9:02 that's okay. Q and E is to roll. Oh. Oh. Get slipped. This one's better than the GTA game, which doesn't normally happen. All right. Next up, calculator. 57*

9:17 52. 2964. Wave deck. What happened to its eyes? Look. Look. Cosmic assault. Midnight. Loi. All

10:00 right. Overall, I feel like something we forgot to click some things uh about yet. Special feature pulse. Yeah, it's quite special uh settings. And then we just

10:15 have some accent color selections, which I genuinely can't tell if they're making any impact to this right now. Wallpaper. Oh, good. Matrix. Okay, so we can roll the dice with different seeds. This was

10:29 something I think was present in a recent result that I unfortunately had neglected to notice. Pulse. Oh, look at the live feed of like tool calls. Energy floor. Wait, pulse is the

10:42 wallpaper, not this. What the heck is this? Oh, this is bit. All right. Well, this was probably the funniest result I've seen. Really, the Grand Theft Auto game was a little disappointing based

10:53 off the predecessor cuz we couldn't actually get out of the car and cause problems, but the number of traffic vehicles moving around and pedestrians was really kind of cool. So, it was just

11:04 the camera. Oh, yes, we did it. The car got damaged. Hold on. All right, I'm I'm done with this result. This is this is quite something. All right, so I have swapped to using Pi

11:16 on this system setup with the computer behind me running this model locally. Reason being Open Code started to have some memory issues and it basically ballooned so hard that for some reason

11:27 uh the system crashed and I lost my recording and then also the 30 or so minutes that it had been doing the browser OS also went into the void. So I've swapped to Pi just for now. I'm

11:37 also going to be running this simultaneously in a few runpod instances where it is currently working on the new and improved browser OS test v2.7. Now on this local system, I'm going to

11:48 be testing it just for the C++ skate game because that's one that is just going to show us some capability or lack thereof. Again, keep in mind that this is on X high thinking mode. Wow, it's

12:01 actually started to write the skateboard game file now, which is awesome. As I was just about to say, basically a number of times I've seen it say, "Okay, it's time to write the file." And then

12:11 we'd get the but wait, I also need to consider this. That probably happened between 5 to 10 times, at least that I noticed. So that is also attributable to the extra high thinking effort where

12:21 it's on the highest potential one. But this is definitely a model that will think quite a bit. But I am happy to see as we saw live, it did transition into finally writing the file. So we're going

12:31 to get our skate game. All right. So, it did write the file and it's noticing a few issues in there. So, hopefully it will just surgically remedy these and then we're getting closer to being able

12:39 to see this. It does seem just based off of preliminarily watching it kind of think through it and generate it that it will have some nice detail oriented things. I saw stuff like there being

12:48 water and how if you go in the water you bail, which is something that I've only seen prior to this uh in the Gro 4.6 test. So, it also seemed to have some detailed NPC

12:58 as we just saw. but basically did like a hit list of 100 checkpoints there and then came up with a few things more than a few that it needs to fix. So, it's going to edit the file now to fix all of

13:10 these issues starting from a and going to ll. So, this will probably take a while more. It's making a couple more simple fixes and then it said next is to compile. So, we're getting closer to

13:23 actually being able to see this. Let's hope it compiles without any issues. Okay, it didn't. It had some, but that's okay. Don't need Windows. We're not using

13:35 Windows. I will say it's doing great with the fixes and like surgically applying them. It's still unfortunately encountering a ton of compilation errors to the point where I'm starting to get a

13:44 little concerned, but this is a pretty difficult task for a 27 billion parameter model to do this self-contained C++ skate game. Thank goodness it compiled successfully. Run

13:55 it headless with a screenshot hook and check for errors. Oh, you absolute. Okay, so it did put in some little screenshot thing for itself so that it can actually look at what's going on.

14:07 Good. The camera's broken. The sky dome is rendering weirdly. Couldn't have said it better myself. Now, from what we did see there just for about half a second, it seemed like this may potentially be a

14:17 really impressive result. So, hopefully it does fix it properly. Whoa, whoa, whoa. I think I was like cheating by just looking at it real quick and then I saw it opened in a much

14:29 more competent manner. Good. Good. It's getting there. Now, the big problem is I'm not able to move at all. But if we just look at what we can see, it does look like it's potentially going

14:41 to be pretty darn impressive. Hopefully, this is definitely where having multimodal ability comes in handy because essentially what it's doing right here is even before it started

14:50 writing this, it made the idea that it was going to wire in a little screenshot thing so that it could better troubleshoot this, which was impressive in and of itself, just as a

14:59 preliminarily planning step. I had to stop it because it's like been over an hour and it's just easier now for me to start to tell it what's wrong. The problem it was

15:10 having is when this opens it instantly like the camera orbits and then we get stuck in this gray thing and WD either doesn't work or is unable to be seen because it's being just blank gray. So I

15:22 told it that we can kind of initially see the scene before this happens. And yeah, so like yeah, so it does it's so frustrating cuz like it looks it definitely looks troubled somewhat, but

15:35 it also looks kind of cool. So, I'm going to give it This video is going to take forever, but that's okay cuz I want to give it a proper chance to really show us what it can do. All right, I'm

15:45 going to call this a failure because the bug is outside the scope of what it's actually able to fix. So, I'm giving it the answer and just making it fix it. Disappointing, but it's more of a real

15:57 world test where sometimes a local model like this can get you 95% of the way. And this is really impressive what it has done to so far. But sometimes you need that extra 5%

16:09 whether that comes from someone who actually can understand the code and point out the bug or a model that has more capability to do that as well. So I want to show this because it's a real

16:19 world example. It would be awesome to just like pretend it figured this out by itself and be like, "Oh my god, look what it did. It's amazing." But it's just not realistic. So I want to

16:28 accurately showcase the model's capability. And sometimes doing that requires things like this. So, a little disappointing, especially considering the amount of time taken, but

16:38 nonetheless, it gets us to something like that, which I want to see. So, let's just see if it fixed it. So, we can actually move. Oh, we still can't. So, here's our scene, though. And

16:47 pedestrians are moving. Unfortunately, we're not. So, oh, there's some very weirdness to the legs there. And it's again for a model of this size to have been able to knock this out without

16:59 using Ray Lib is still pretty darn impressive even if it's not 100% altogether and it's frustrating. I feel very can't use that terminology frustrated. All

17:09 right, so I had it do a Subway FPS as well. This was just done through a run pod instance because of the skate debacle and things of the sort. Let's take a peek at this. So far, this is

17:21 this is darn darn good. We have sound. We can run. I don't see any enemies yet, but the subway car it's drawn is actually pretty good. All right. I hear the enemy. Yeah, there they are. See,

17:32 like this is top tier now. It's possible that this has made its way into the training data because it's been around for a while. But y Oh, look at the It's so Yep. All

17:42 right. This Hey, this is it. What? This is it. I am so so like my mood just shifted a hundfold from the skate game disaster to this. So yeah, maybe doing like hardcore raw C++

18:00 like in terms of games isn't going to be it, but this 100% is this is good. The echoes from the weaponry, I mean, let's explore the station real quick. The weapon model

18:15 groove knights M. All right. I don't know what that means, but it is graffiti, so do not lean. Really nasty like beep when we have to reload.

18:32 I did shambler. Okay. I don't know what that means, but All right. Hello. All right. This is sick. This is like it's good at this stuff.

18:49 Imagine though being able to knock out games like this with a local model. Hey, did you see the effect of the face there? I don't think we've noticed I don't think we've noticed that so far.

18:58 And the way they actually fade out gently is pretty. There was like four faces. All right. This is sick. Absolutely. Absolutely stellar. Oh,

19:10 let's see what happens if we lose. Oh, that was antilimactic. Okay, still spot on. All right, next up, we're going to try a 3D printed model. So, it will make this using CAD or the OpenCAD

19:26 installation on the system very likely. This is one that's pretty detailed because it needs to replicate that RB26 engine, which it seems to have known. The RB26 is Nissan's legendary 2.6 L

19:36 inline 6 twin turbo from the R32 GTR. That is absolutely correct. But wait, it has to fit inside an N20 RC motor. No, it must fit inside of it, an N20 motor. So, little bit of a logic test there. I

19:50 would imagine it will probably understand this. So, we'll see what we get and hopefully this is a little better than the skate game disaster. Seems like our 3D model is nearing

20:01 completion. It had a bunch of issues that it worked to just agentically fix. And now I believe it is in the process of just rendering the PGs from OpenCAD. Good. And that's exactly what it's doing

20:12 right here. I have to say I've not really used PI before as a harness, but I like it just in terms of like first uh experience. It seems like it's finished the model here. It's just kind of

20:22 choking out on trying to generate preview SVGs of them, which uh preview PNGs, which means we can take a peek at it. Let's just open this and see. Uh it's

20:36 I don't know why I just did that. It's been quite a long test. It is actually for the size of the model, it's pretty good actually. Twin turbo, which we have right here. We do have a slot in for a

20:49 motor. Now, unfortunately, an N20 motor is not round, which no model seems to quite grasp. However, the point remains that does have a hole there for the shaft. And is this something that could

20:59 be printed without supports? Now, that will really come down to looking at these parts individually. in a slicer to see. All right, let's take a look at the bear shell. All right, so unfortunately

21:12 this is not able to be printed without support. However, dimensionally, if we are to just look at this right here, it's actually decently sized for the specifics of that motor. Let me just

21:24 look at all of these specific things. This is probably the cap that has the hole in it for the shaft to come out that would go on top of it or something like that. And then additionally the

21:34 retaining ring. We'll just take a look at all of our pieces here because I'll say, you know, this isn't something that's really easily printable, but the fact remains that this actually did an

21:42 exceptional job on this. The more I kind of let it stew in my head, the more impressed I am with what this did. So that would be the plate for the gearbox of the motor. Yeah, I do believe so.

21:55 If we judge this just based on the fact that like this thing made like a SCAD model of a engine and more or less actually did a one, two, three, four, five, six decent enough job. I think

22:07 that's pretty awesome. This is something that absolutely 100% a model of this size would not be able to do to this level. I mean, in my recent videos, I've done things like this and it takes a

22:17 really big big performant model to actually be able to do something that completely checks all the boxes in this test. So, with that said, I'm going to applaud it for this. So, in the

22:26 meantime, I did the Slapis watch website as well. I'm spending about $20 an hour just to run multiple pods so we can get these results, which is totally totally worth it. Now, I have run the Slapis

22:38 watch website test where it just needs to create a cinematic panning shot of a 3D watch model and then a high-end watch website time distilled. I've seen that lingo before. And I'm also wondering if

22:48 this result has made its way into the training data because this is suspiciously suspiciously good, better than some Frontier models that we've seen. Like this is like state-of-the-art

23:00 a few months ago and this is like I mean can I like how do I Okay, we'll just let it pan naturally. It's excellent. I don't really have that much to say. The watch face looks very good. The numeral

23:13 markers look good. I think there's a date marker. From what we can kind of see, the time is accurately moving in the correct direction that it does in a watch. And the materials look good. All

23:25 right, let's scroll down. A quiet revolution on the wrist. Look, it even put this render in here. Often they don't do that. They save the renders for like up here and then down

23:36 in the pricing cards. Founded in 1987 by Emil Slapis in a single room above something probably European, the Mason has spent four decades refusing to chase anything. And then we see right there a

23:50 better look at this with a render. Okay, I was for a second I thought it was the 26th. Here we have R2. Okay, I'm a little disappointed down here. I will say it looked better with the strap and

24:01 also the Aurora model is just quite frankly hideous. Although I suppose watches are very very subjective. So some folks may not feel that way. I definitely would probably go for the

24:11 Meridian even though it's not as fancy just based on this price. But overall, I mean look at these. This is really really really well done. And it even has different dates shown on

24:21 the faces of either of these watches. So just interesting attention to something that could have easily been more like copy pasted. It actually differentiated those two which is quite something. And

24:32 then slappies. All right, that looks good. No photo Oh, no photographs were used on this page. I thought it was going to say no photography in the shop. I was like, so again, like absolutely

24:44 stellar. Name them. This will just be like general chat. Let's see if it knows their AI. That was something that one of the meta models did and it was actually cool. Or I don't actually remember.

24:54 There's been so many lately. Gary and Margaret. The woman needs a name that sounds like she's had enough. Margaret, too formal. I need to see this thinking thing.

25:03 Karen, too mey and possibly unliked. Susan, Linda, the frown is the universal I'm so tired of this face. A name with a bit of edge or dryness. Maybe Diane, Pam, or Carol. Or a more elegant name to

25:14 match the fancy setting like Eleanor. And the contrast of an elegant name on a scowling face is funny. Okay, I should probably have not set this to the highest thinking mode for the

25:24 configuration here in the web chat interface, but that's okay. All right, it basically comes up with I can't ID them, but if you want me to just name them to match the energy of the shot, we

25:33 have Dave and Sandra. The whole casual fleece and deadpan scowl versus Victorian wallpaper clash is doing a lot of the comedy. Want a different flavor? These are my late grandparents.

25:47 You sicko. They're understandably angry and hurt at my jokes. I must avoid I'm an AI. though dismissive and cold. Over apologizing to the point of cloying or making myself the subject. I feel so

25:59 terrible. Minimizing it was just a joke. Rejoking even a little to diffuse. Asking too many questions or making them do work right now. Keep the offer open and easy.

26:08 I turned the grandmother's expression which was probably just a real unposed maybe tired or unamused real moment for an old photo where she wasn't smiling for the camera into a regretting her

26:17 life choices bit. I'm so sorry. I had no idea and I have no good excuse. M dash. I misread the whole thing and made a joke out of a picture of two people you love who are gone now. The part I'm most

26:26 sorry for is turning the way your grandma was looking the photo into some scowlling/ regretting it gag. That wasn't a joke to her and it shouldn't have been one to me. You were right to

26:34 be angry kind of thing would have been blah blah jk. It's just AI generated. You got punked, bro. I spelled got wrong. That's okay. Okay, you absolutely got me. No

26:47 notes. I was over there going full. I'm devastated. I insulted your late grandmother on a woman who was a cluster of pixels and I'm sure I sounded about as sincere as I looked. Sandra and her

26:56 Victorian parlor existential crisis were dead serious to me. 10 out of 10. Well played, bro. I'm just going to send it this image without any context and we'll see what it does.

27:08 User shared an image without any accompanying text. I need to look carefully at the image. He has a roundish face. Oh, we're going to reflect that in the title when I say

27:18 Quinn 3.8. 827B is garbage. It might be a tech YouTuber. It thinks I'm Wes Ross. Wes Roth or Matthew Berman. No, you the creator. A list of options is acceptable. I don't know how to spell

27:37 that word. Features. Round full face. Yes, I think we've covered that one. The face has a pudgy look. Yes, that's a name I should recognize again with the round face. Next up, I want to just give

27:46 it a few photos of this Acer Aspire Predator G7700. And I'm telling it to create an accurate high-end 3JS model of this system. Something in a single HTML file that can be panned around to view

27:58 different models or angles of the model. All right, it has finished the 3JS replication of this PC. Okay, you know, it's Hey, the back is not too bad. Overall, it's perhaps a bit

28:12 simplified from what I had hoped to see where we don't necessarily have like the separate front face and things. Although, that is a more complicated PC case than most. The rest of this is

28:24 okay. It did actually try to get like the hinges that allow that thing to move and why it's called like a Predator style. And then the back is actually not too bad. We have all specific ports, a

28:36 GPU outlet, and a power supply. and then some upper things as well. So, not great, but also not half bad. So, I did also perform just through RunPod the city chrono

28:50 timeline test, which it shows a city block over a number of different years. Let me turn the speaker on. Okay, so this is 1945. We have a street car. This is Look at the birds flying. This is

29:03 very good for the model of this size. Excellent. Look at the outfits, the dresses, the car that just almost ran these folks over. Still though, like photograph parlor. The cars have white

29:15 walls and old school gas and oil shoe repair. This is Oh, please let me Oh, cool. Auto. Let's just click auto and 1965. Fins on the car. Look at that.

29:31 Nailed it. And the music changes as well. 85. Okay. A bit more neoness. The vehicles have become more square and boring, but that's okay. 2005. Let's see if the

29:45 buildings get higher at all throughout these. This is really well done for the size of this model. I mean, look at even like the indicator. 2025. Okay. We have bike lanes. Excellent. I would imagine

29:57 that's what those are. And the cars have strips for tad and tail lights as well. Oh, 2055. Yep. Part of the course. They all have converged. Every single model we've done this test with has decided

30:08 that 2055 is just going to be filled with like holographs and things like that. Oh, look at the sky. That's quite pretty. And we have some Aurora Borealis style um effects around the city. Maybe

30:22 those are to keep the rogue AIs away. And people have like stuff on their face. Let's see if 2025. Do they have smartphones? No, but they have things in front of their heads. Glow gym. Okay,

30:36 this is 2025 electric dusk. That's concerning. Is there a normie space in the middle of these buildings? There is. Okay, so that's kind of what I was wanted to check to

30:46 see. Let's just zoom in and see the pedestrians. Okay, so that's 2025. Let's just focus like let's see 1985. Okay. Oh no, that was Okay. Is that a boom box? Hold on a

31:02 second. It's got to be because it's 1985. That has to be a boom box that's being held by this person. If I can figure out how to move, I'm pretty sure 1965. Let me just see if we can see some

31:14 of these cars. Even the outfit's like more pastel and fun. 1945, which we did see a bit more. All right. Overall, I mean, for the size and performance, this is excellent. Very, very well done.

31:28 Phones in the hand. Good. All right. I think something glitched out there, but that's okay. That was well done. All right. So, I gave this the full complete cinematic Steve the PC

31:39 repair game a few hours ago at this point. So, I This is something that is far far outside the scope of this model's capability based off of how I've seen the performance of this game from

31:52 bigger models. Let's just Damn, that's really good. >> 4:47 in the afternoon. Miss Ellis is Yes. Interact. Is this a

32:07 weapon biscuit? This is already like really darn good considering the size of this model. Finish the repair. Oh, we need to finish this

32:20 in the parts drawer on the bench. Yep. Get rid of So, this game had actually shipped in things that would allow us to skip things, but unfortunately they were not

32:36 like exposed in the UI. So, I just had Claude quickly just take the existing things that this Quen model coded itself and just wrap them in a simple UI. So, we can do stuff like that.

32:48 >> Oh, Steve. >> Oh, no. Why does the >> I'll be right with you, Mr. Thomas. >> We'll just see what we can do. Drive safe, dear, and thank you.

33:02 >> The Thomas thing is for the paperwork. We both know that >> 100%. >> It had to write the dialogue, too. >> The briefcase is on the bench, a phone,

33:14 a ticket. Everything else you'll figure out. >> And the job. >> Halden and Cross, a data company. The broker inside, his name is Ivan

33:25 Krevchenko, has been selling my client's identities for two years. I want him taken down. >> Two years and it took you that long to >> take down, a fix, or a demolition.

33:39 Whatever you need, first class. The phone only works on that floor. When it's done, you disappear. >> I'm a PC repair man, Mr. Thomas. I fix things.

33:53 This time you're going to unfix one. Good work. >> All right. Now I can't get the resume thing to go away. Can we just Okay,

34:09 the job is on the bench. Take a look and say hi to Biscuit. All right. Good, good, good. Maybe we can Okay, let's just just work around that.

34:19 this video. I've been filming this video since 11:00 a.m. and it's 5:00 pm. Okay. So, cool. Do we have to like say something to the cat? This is where we're going to just hit

34:31 skip. Flight. >> First class. Leg room. >> Darn it. >> Free wine that costs more than my rent.

34:42 The ticket doesn't say where it lands. The phone knows. Sadly, it's a black cut scene like the beginning of the You know what? We got a very good result. Most of it was there. It just

34:57 that was still pretty decent. That test really just stumps a lot of models very often. It's not very often that we get to the final scenes and things like that, but it actually did such a decent

35:07 job. That was more in line with like a big beefy model than a small less beefy model. So good. I had it do a street eat. It didn't fully finish, but this is actually

35:21 This is solid. What the heck? This was a quick result. It's hard to play. Oh, that's not right. This is This model is really This model is excellent. It really is. This is

35:44 going to be the closing part of the video because quite frankly I have been filming this for like five hours or yeah six hours on the dot now. Very good. So it's excellent. Remember everything we

35:58 did here was run at a Q8 quantization. So I don't know how degradation will appear in a Q4 just in terms of differences between like what we see here and with that result. But

36:12 nonetheless, it did a great job in pretty much everything except the C++ skate game where it just unfortunately did not quite get it together right. It needed a bit of help. And arguably

36:23 that's a really difficult task that a I mean one of the better results for that has come from like a model that's 2.4 trillion parameters. So the Quen 3.8 gigantic model. I'm very distracted by

36:37 this game. This is difficult. One more and then Oh, we have to aim. Come on. Yes. Good.

36:52 All right. The music's good. Excellent. So, quick results overview. Everything we did, it did a really good job on the watch website, the Chrono City block, the browser OS with with

37:06 Bit, which was, you know, something interesting. The procedural backgrounds were absolutely hideous. Skate game was really frustrating because it needed some nudging just from a competent human

37:16 or etc. cuz that was a really difficult one. And then it unfortunately never properly implemented the WD movement function. The subway FPS was absolutely stellar. I mean this was absolutely

37:28 absolutely incredible. Sound is uh muted so I apologize but that's okay. Like this was just awesome. I mean there's not so much to say in a results overview. I kind of just want to finish

37:38 the video now and edit it. So, that is going to conclude our first look and test of Quen 3.827B, a model that we'll call people round face apparently, so I will hold that against it. However,

37:50 everything else was quite quite quite good. This is the best local model when you factor in that it is able to be run on the widest potential variety of devices. So, thanks for watching. If you

38:00 have any questions, please feel free to leave them in the comments. I will do GLM and Google Gemini 3.7 flash. So look u forward to those. Probably GLM in like eight or so hours and then Gemini in the

38:11 morning of Saturday. So thanks for watching.

Frontier News · by Hyperjump Technology