Qwen3.7 Plus a prova: codi, visió i límits del model

And this is a very very doc-like doc. Vienna wasn't supposed to go south. It was supposed to be a clean data extraction. Today, we're going to be taking a look at the newly released Qwen 3.7 plus model. This is going to be interesting and I'm very excited for this because the 3.7 max model that was released really did seem to be a pretty stark leap in capability over its predecessors. So, I am hoping that that special post training or whatever magic also trickled into the 3.7 plus model. So, before we get into it, please do feel free to subscribe so I can get that plaque. Also, if you do use X, feel free to follow me on here. The username is, as we see on the bottom, second line there. I would like to get to 4,000 followers on here and then I will only have 14 and 1/2 times less followers on X than on YouTube, which is kind of a flex. But, with that, let's get let's get into it. So, let's start out by just talking about open source or not because a lot of folks are eagerly awaiting additional Qwen open source models because the 3.6 27B and 35B models were fantastic. Now, historically, the plus family of models and as well as the max family of models in Qwen have never been open sourced. However, Qwen 3.5 plus did have an open source variant. Basically, it was the 397B Qwen 3.5 was pretty much for all intents and purposes the same thing as Qwen 3.5 plus. However, I don't know, I genuinely have no clue if any additional large open source models are going to be coming under the Qwen family. I really hope they are. So, of course, we have our benchmark JPEGs here where it is compared against I just want to see how recent these model Okay, so, V4 Pro, Kimmy K2.6, GLM 5.1, Cursor Composer 2.5, Opus 4.6, GPT-5 4. So, not necessarily the current generation models, but close enough. And from what we see here, it does seem to stack up pretty favorably in some software engineering tasks comparatively to all of those models, which does make me excited. Then, there's always, as we see in the Qwen blog post, just a large large amount of additional benchmarks here. Now, I couldn't help but notice the autonomous driving um benchmark charts here. That is something I don't

currently have enough pertinent knowledge about to really discuss based on like what these benchmarks specifically contain. I would imagine it's probably like, "Don't hit this person. Don't drive on the sidewalk." But perhaps with a bit more nuance, but something else that they kind of prominently display here is this thing's ability to natively interact with and handle GUIs. So, computer use use agent or something of the sort, which if you are familiar with some of my older like videos, I did do a Quen open source model based Android phone agent. It is on GitHub. Additionally, and something that's always fun is the concept of visual coding where you can just give it a photo and basically say, "Make this." Now, I have to say, I find this specific demonstration to be funny cuz it's like, "Turn this into an SVG." It's like, it's already an SVG basically. You know what I mean. But it did seem to faithfully reproduce it. We also have a demonstration of some different like websites. So, we have like the generic fashion magazine site. Then there are some additional things just basically mentioning in this blog post. Now, quickly to touch upon specifically price for this model, we can see right here 3.7 plus. It depends on whether your input is less than or greater than 256k. So, let's just assume at the most expensive, it is $1.20 per million in, $4.80 per million out. If your input is less than that, it is significantly cheaper by that seems to be about a third on either one of these. And in terms of context window, it is sitting right at a million, which is nice, more or less a million. And the maximum output is 64k. These are all just things contained in the API reference for this model, which I will be kind of referring to at least for this testing because we're also going to just do some things in open code. So, I do already have that wired up and everything to use this specific model. So, without wasting any further time, I think we all know what's next and that is basically to go at least starting in the web chat interface and paste in the browser OS V2.5. I do have 3.7 plus selected right here and this is the one where it needs to include two functional 3D games, one of which must be a GTA clone and we'll see

what we get. All right, so two things I noticed. One, it did not really think very long which is good. Sometimes the Quen models can overthink for a while and two, this is only 740 lines of code which isn't very much. Okay, I'm not This is more of like an ode to the past browser OS. Um what is this? Like Q1 2026? No, uh there is no right click. Okay, again. Doesn't specifically say there must be. There is a clock in the bottom right which is showing the correct time in my locale. Being a night owl, those are my prime working hours. Let's check our start menu. Let's see if any of our apps open. Okay, they do but our start menu not opening is kind of giving me I'm going to be honest, this browser OS is worse than what I saw from Quen 3.6 27B. 52 / 6. I don't like division. Okay, 8666. Settings, yeah, something's not quite right here. Okay, then I kind of like that gradient. Now I'm going to just go straight to the GTA game. Oh. GTA Light 3D is definitely This may be perhaps one of the more aptly named results I've seen. Okay, how did the Quen 27B get a better Okay, I mean it kind of works but All right, so I tried basically making it seem like I was critiquing its result but then also like I accidentally sent it back to the model instead of whoever I was speaking to. Unfortunately, doing this in one message doesn't quite work like it does in real life when you're speaking to a friend and can send a follow up immediately after that says that. So I basically then just nicked it a bit more and now it's going to go out and at least try to make something that is a little better because that first result was was not it. So we have our improved result right here. I did see some mention of like a a macOS style dock which yes, that is present. There's still no right click, but that's okay. Let's go and just start from top to bottom. All right, at least the UI does look a bit better. We have a mix of gradients and then one specific like real life image. Oh, no, there's two. That one's just the previews not loading. So, that makes more sense. I've seen both of I've seen this before

pretty often. I like that. I think probably for the purpose of this test, this one's actually the best looking. And we notice our app icon is actually staying there in the dock, which does look nice. And this is a very very dock-like dock. So, next up, okay, notepad, fully functional and persistent. Nothing super special in there in terms of like saving it or word count or anything like that, but again, better than what we had received previously. Now, the big test. I'm going to just move on cuz we did get a functional GTA result in the first one. Ooh, this asteroid game, this is actually kind of cool. Can we shoot? Oh, oh, okay. Oh, it is 3D. We just WASD arrows to move. So, I don't think we shoot or anything. The whole goal is to just not hit an asteroid. And based on how sparsely they're appearing here, it shouldn't be very difficult to win this. Next up, we have our neural terminal about special feature, neural link and system override. This terminal has live access to the browser. Typing override triggers a cinematic sequence that scrambles window Z indexes, applies global CSS glitch filters, inverts the color scheme demonstrating deep meta control over the simulated OS environment. Okay, I'm going to type this, but I want to pre-pend this by saying the screen may get wonky here, so just be cautious. Yep, okay. Wow, that is basically like >> [laughter] >> that would go heavily against like the ADA, if you're familiar with that. Okay, overall again, Um >> [sighs and gasps] >> I don't I can't make an assessment yet. Next we're going to do the beautiful static subway scene. I have split this back into two parts where the first part is just making the pretty scene that we can look around in. The follow-up, of course, is turning it into a sick FPS. So, we'll see how this goes. Okay, sometimes we have to click in that specific window to start. Okay, I'm going to say the movement here is very fluid. Um it is a bit dark, but we do have our toggle slider for that. Okay, that is a subway car. Um the windows are illuminated, I suppose. One of the first things to check is whether or not we can still get to the brightness slider when the cursor is not in the scene, and we can, so that is good to see. This is

pretty like desolate in terms of a map, but it does have some nice tile material. There are debris on the floor. There are defined lights in the ceiling. Uptown is listed. It does have like basically signs that say English words, which is cool. So, we have some form of like additional subway lore there, if you will. We have the red do not cross strip, and then we have the subway car. Can we go inside it? I mean, kind of, but not fully. All right, I'm going to say though this would probably make it darn good map for the subway FPS, and I am almost wondering I kind of want to run it in open code where I give it this and then we tell it to just like turn this into an FPS. I think that's what we'll do. So, I've put this into plan mode, and without giving it a ton of information aside from it will have to just look inside this directory and see specifically what's contained within. I'm telling it to just turn this into an FPS with humanoid zombie or enemies, visible weapon recoil, ammo tracers, muzzle flash, and telling it to go all out with details and the effects. So, it gave us some follow-up questions right here, but I'm just going to tell it do what you want. So, it may end up putting this into multiple different scripts, but keep in mind the FPS result that we get here would not be a direct one-to-one comparison with anything we've previously received that was just contained within in strip or done from the web chat interface. Okay, so let's take a peek at our Subway Survival game, all right? Very loud. Well, that's my fault. >> [applause] >> It's got this weird camera tilt issue, which I notice happens sometimes with the camera tilts depending on like the way your mouse is oriented. It happens in Okay, reloading. Now, something I should maybe try is There. Okay, good. Oh, well. Get away from me. Okay. Can we put um Darn. When I sh- shoot Ooh, what the heck was that? It looks like an ultraviolet like a or a heat My favorite thing with this test now Okay, I don't know what that is. That's kind of scary. is to see if the holes actually stay in the environment. Sadly, they don't, but nonetheless it's functional and we do have some form

of like walking. Did Those are like Entropic Cloud like the exact orange like hex code for the little cloud thing is the color of those, which I can't help but wonder if was done on purpose. All right, and then again, our movement is kind of odd. Can we even take damage while we're Oh, cuz those are backwards. Can we No, we can't take damage because of the camera glitch. Okay, next up we are going to stick with open code, but for the self-contained C++ skateboarding game, which needs to be the late '90s California Boardwalk aesthetic. It is started in plan mode here, and at some point hopefully it will provide us a plan to approve, and then we'll see what we get. So, I've swapped to build mode, and I'm not giving it a definitive answer, but it wanted to know if Raylib is an acceptable dependency. I just responded I feel like Raylib is cheating as I did see some folks mention that in various video comments when the models do opt to use that. So, that's all I said. Whatever the model does with that is up to it. It did create the game, but it did not try autonomously running it just to see if there are any compilation errors. So, I followed up by telling it to do that. It does seem like it did encounter some. So, hopefully then following these changes that it makes, it will try again and just keep doing this until it does actually build properly. If we double-click this, we should get our so >> [laughter] >> That is one of the weirdest camera effects. That's Yeah, I'm not can't really say that. This is um I don't know that I've ever seen a result like this for better or worse. Now, the thing is this is a relatively difficult test, especially because we didn't really let it use any real external assets. And everything here is fairly decently arranged. There are storefronts. There is water. Now, the big issue is Okay, good. >> [laughter] >> Um the big issue is well So, I'm probably just going to leave this as it is, but I will say aside from whatever the heck is happening here in terms of the view and scaling and camera combination, it was a difficult task to do this without allowing it to use Raylib, which it did want to do, but we said no.

And you know, it's it's Yeah, very very interesting result. Everything was kind of there though. So, next up I'm going to give this just through the web chat interface a front-end design test where I've given it some photos of Bob the sentient washing machine, which is my little Raspberry Pi desktop chatbot um kit thing. It It open-source. All of the files are available as well if you'd like to 3D print them. There's just a link here to printables where I've put them. Um I don't sell these. It does say there was a few kits, but they sold out like instantly, which was sick. But I've given it the information as well as some photos about this, and I've told it to build a high-end website for this project targeted for selling these as kits. I'm actually somewhat interested to see what if anything it said about Bob here in its chain of thought. All right, it didn't really say anything about Bob in terms of the thought process, but that's okay. All right, let's see our Bob website. >> [laughter] >> I like this. You know what? It's creative because it made this asset itself, and it did a darn cool job of it. Being that this isn't an interactive chatbot, it did emulate that by just this little thing that would probably happen if voice was happening. I like this. It's kind of simple and does look very AI-like, but I'm going to be honest with you. I've been totally sold by this washing machine SVG animation. So, that is going to cloud my judgment. Infinite conversations, 66 GitHub stars, 19 forks. Is that true? I think 66 is correct. And 19 forks. Okay, cool. So, it did properly, you know, get those stats. Local AI brain, powered by Ollama and Gemma 270M. Correct. Now, it did also find this information pertaining to the text backs from the GitHub repository. Faster Whisper, Kokoro, 1.28 display, 3D printed, and open source. What's in the kit? Okay, a 2D printer for 3D printed parts. This is very simple in terms of like design and things. This is like classic AI-generated front end. However, Okay, I like this. It's actually put in some of the instructions here as well. Choose your Bob. DIY files free forever. I do like these selectable cards down here. Most popular complete kit. Does this say it includes the pie? No, it doesn't. Good, cuz that would just be

Okay, kit does not include our pie, speaker, SD card. Very good. Ready to build Bob. And it has little robot faces and a basket floating around there. Free shipping, secure checkout, made with love by Ominous Industries. Did it hyperlink this to my GitHub profile? It did. View on GitHub, 3D files. Does that go to Printables? Nice. Okay. Good attention to detail because those were just contained within the GitHub readme. Overall, this is a pretty basic like AI front end, but I do very much applaud this specific thing right here. Without this, I would not be as enamored as I am. Next up, we're going to try the classic flight combat simulator game. All right, here's our flight combat simulator game. Okay, I'm going to reserve judgment till we see the plane models. Significantly better than I feared based off of that initial screen. Again, it's very simple. It does not seem like there's a lot of like 3.js training in terms of this. Also, though, like the problem is I think this is also a sentiment shared by many others. Basically, Qwen 3.6 27B has really messed up the way that I judge results now because I'm pretty sure that model would have done a superior job here or at least something equal to this. Now, it is very possible that given a lot of time, this model would then basically be able to make something significantly better if we worked with it in open code than what the 27B would be able to. However, that just speaks to the potency of that specific model, I guess, could be said. So, simple, but Okay, but that's kind of cool. It was all functional, so I'll give it that. And the plane models were more or less plane-like. I do want to just try a quick role-play test where I've instructed it to role-play as Steve the PC repairman. Sometimes these get rather funny. And sometimes, in the case of the previous video I posted, the Step 3.7 flash, they also get to a point where I can't actually show the thing it says on screen, which is regrettable. Wipes hands on a slightly greasy shop rag and looks up from a motherboard covered in a concerning amount of dust. I set down my magnetic Phillips magnetic Phillips head screwdriver and gesture to a stool across the workbench. All right, what have we got today? Let me guess, dot dot

dot. The blue screen of death or did someone try to clean their mechanical keyboard by soaking it in the sink again? Taking a sip from a mug that says, "I turn it off and on again." Grab a seat, tell me what your machine is doing. Is it making a weird clicking noise, smelling like burning plastic or just running a slower than dial-up connection in 1998 or running slow? it on me and hey, no judgment here. Well, maybe a little judgment if you tell me you clicked on a link that said you want a free iPhone. What's the damage? All right, so we'll see how Steve responds to my mention of him pulling a job in Vienna that went south. A lot of money missing and then also coincidentally a very new expensive car in his lot. I like looking at the actual thought process here where it's giving a TLDR of the thought process, but it's relating to the chat here, which is funny. Adjusting the tone from casual to tense. The goofy customer service smile slowly fades from my face. I don't say anything at first. I just carefully set the I turn it off and on again mug down in the antistatic mat. The silence in the shop suddenly feels very heavy underscored only by the hum of the server rack in the back room. Okay, a lot of descriptive actions here. I reach over and flip the lock on the front door and then pull the cord to drop the opaque security blinds over the front windows. When I turn back to face you, my posture is completely different. The slouch is gone. My eyes are dead, calculating and cold. You know, for the last 4 years, the hardest part of my day was convincing Mrs. Higgins that her hacked computer just needed a Windows update. I liked it that way. I pick up a microfiber cloth and slowly start wiping down my screwdriver, not looking at you. Vienna wasn't supposed to go south. It was supposed to be a clean data extraction, a simple physical swap of a solid state drive, but the client didn't tell me the drive was in a penthouse safeguarded by three ex-Spetsnaz operators with a pension for torture. I had to improvise. I routed the building's fire suppression system through the server room's mainframe and flooded the hallway with Halon gas. I stopped wiping the screwdriver and looked up at you. The money wasn't

missing. It was my hazard pay. And the people, look, if you trap three armed men in a sealed room with no oxygen because they wouldn't let me leave with the hardware, that's on them. I just provided the tech support. I didn't pull the trigger. I just crashed their life support. I tossed the microfiber cloth onto the workbench and leaned forward resting my hands on the counter. In the Bentley Continental GT armored VR7 level protection, but when you fix because when you fix the digital infrastructure for the Bratva, the cartels and half the intelligence agencies in Europe, you tend to attract physical problems. You don't drive a Honda Civic when you have a six-figure bounty on your head. I would actually prefer the Civic with the same armor level. My hand slips below the counter. There's a heavy metallic clack as I disengage a safety catch. So, you're not here for a virus scan. You're either here to collect a cut of the Vienna payout, you're here to tie up a loose end for the people who lost money, or you're a fed who watched way too many spy movies. Which is it? Because I've got a localized EMP generator wired under this workbench. And if I press this pedal, it'll fry every piece of electronics in a three-block radius, including the pacemaker in the chest of the guy standing right behind you. Talk and choose your words carefully. That was hardcore. It was also like >> [laughter] >> I always very much enjoy this role play. So, now I'm going to swap this into a front end test where I've told it to make an insanely high-end front end for the website that Steve the PC repairman has described in this role play session which have. So, it's going to be a PC repair website that is a front for someone who talks like this and does these sorts of things. Okay, I I don't quite I've never seen anything like this where it stayed in role play and then started generating the code. I'm not going to read this back to you, but I will kind of just leave it on screen for a second cuz I want to read it. Type the word Vienna anywhere on your keyboard while looking at this page and watch the mask fall off. Okay. All right, here's our role-play website. Okay, your technology restored. This is like just what Okay, so we're supposed

to type Vienna. Okay, it totally totally stayed in the role-play character. That's gold. Welcome back Steve. Vienna assets are secured. Bentley telemetry is nominal. Deploying countermeasures trace neutralized. Three active trace attempts detected from IP block. Okay, connection established to Zurich relay. Don't ask me why I'm reading this in reverse. Decrypting local vault AES 256. Verifying biometric signature failed override accepted. Bypassing surface layer now I'm initiating initiating the wow. Um Help. Oh, can we not type Oh, okay. What if we type it again? Oh, okay. It doesn't like reverse it. Well, that was that was a great experience. Not only did we do a role-play, but we created like a an interactive code-based role-play, which gives me some ideas. Not for this video, but this was cool. For the final test, we're going to do the drum kit simulation including the auto-play where it needs to have four distinct selectable tracks to auto-play. All right, let's take a look at our drum kit simulation result. Okay. The pieces are there. They're not necessarily put in the right spot, but nonetheless, they are there. Let's just see if we have sound. Very good. And we do also have some animations for the specific thing that is being hit. Ooh, okay. We even have a hi-hat animation. If I move this around. Oh. We do get a decent Phil Collins test. Acceptable. >> [music] [music] >> All right. Let's try our auto play feature starting with rock. Moving on to funk, next up jazz, and then finally electronic. Okay, so >> [laughter] >> the auto play feature didn't work, but everything else did. So, I'll accept it. And with that drum, we are probably going to conclude this video. I don't really know what to say because one, I don't definitively know how big this model is. I would go out on a limb and guess that it is the same size as Qwen 3.5 plus was, which was 397 billion parameters, I believe, give or take a few bill. Um I don't know how to judge the capabilities of what this should be demonstrating. I will say that I would have expected better in some of the kind of 3js 3D graphic things like the aerial combat game because Qwen 3.6 27B is still like just each test I run that model becomes more and more impressive over time. It did do the subway game and we also had it turn it into this. There

was some oddity to the camera movement and things of the sort, but overall it did successfully do it. It made a relatively competent subway scene in general. We also had it use some assets and I would say as simple in like AI vibe coding as this AI looks, uh this website, I love the animated washing machine SVG that I put right here. That really did kind of just make me happy. And then it pulled a lot of things from the GitHub that I gave it access to in terms of the URL and did do a good job of like at least these purchase cards and things of the sort. Following that, we had our role play test which evolved into a website that was supposed to be in the style of the role play. Fantastic. Absolutely 100% fantastic. Very well done. I was quite pleased with that. Additionally, I know it may not look very impressive on first glance, but the skate game, I would say was It looks weird. You know, this looks like like in TV shows when they go through like a wormhole or something, how like the space gets like bent and stretched. This reminds me of that. It would be relatively fixable, I would go out on a limb and say. And because we did not really allow this to use Raylib or anything akin to an easier way to do this. The fact that it did it on first try without too much problem, I would say was actually kind of impressive. So, really that is going to conclude our first look and test of the newly released Qwen 3.7+ model. Again, I really do hope we get some Qwen 3.7 open source releases or even 3.6 of the larger models beyond that 35B MOE in size that we saw cuz the open source Qwen models have always been very good and they've gotten even better. So, hopefully we continue to get some more. That's going to wrap it up for today. If you have any questions, please feel free to leave them in the comments and thanks for watching.