LLM Battle 2 - Qwen, Ornith, Qwythos, Gemma 4

[00:00:00]Part two in what I hope will become a long-running series. In today's challenge, we're going to pit a bunch of local LLMs against each other to see which one can come up with the best answer to a given challenge prompt.

[00:00:09]We're going to have a qualifying round to narrow the field of competitors down to just two. Two will advance to the final round where they will be given a more difficult challenge to solve. The results will be judged by me using a completely arbitrary and non-scientific system of judgment. Can Quinn 3.627B retain its championship title? Let's find out. [music] Local host.

[00:00:30]No place like local house.

[00:00:33]>> Okay, qualifying round. We've got four contenders today. We've got Quen, Ornith, Quithos, and Gemma 4. Uh we're going to do a quick qualifying round to narrow that down from four to two. Uh the the name of the game for the qualifying round is Asteroids. We're going to do a clone of the classic arcade game Asteroids. This should be a fairly easy one, I think. Uh I've laid out a few constraints graphics wise.

[00:00:54]We're just going to have a fixed size window. Uh, I describe uh what the ship looks like. It's a triangle. It's not that hard. Uh, the ship can fire projectiles. Projectiles can break apart asteroids into smaller asteroids. You've got to shoot all the asteroids to win.

[00:01:05]The asteroids tumble around randomly and they can pass through each other. The asteroids don't impact each other, but they can impact the ship. Uh, collisions are described. Uh, the output format is a single Python file. We're going to be using the Python uh, the PI game library for this. No other third party dependencies. This should be a standalone uh, thing that you can run with one command. Uh graphics are pretty basic, but I'm going to make a couple of notes here. I I want the asteroids to have like an irregular shape. They should look like rocks. They shouldn't just be simple circles.

[00:01:33]And when an asteroid is hit, it should have a little explosion with colorful particle effects. And uh I also kind of want to see when you're applying your thrusters, you should see those thrusters actually firing. Uh and then we've got a creative bonus feature. Uh this is I want to test out the imagination and the creativity of the model in question. So I I'm going to say add one bonus feature of your choosing that enhances gameplay. Uh you hit B to toggle the bonus feature. Uh the key point of the bonus feature is that it should deliberately violate at least one of the other constraints. And I do list a few examples here like being able to shoot more than three projectiles at a time or disabling friendly fire or enabling collision detection between asteroids. That might be fun. you know, like uh there's different things that the LLM can do. These are just suggestions. Uh you're not limited to this list. So, I want the LLM to pick something and surprise me. And I specifically end by saying, "Try to be creative and make it fun." That's that's deliberately vague. What does fun mean?

[00:02:30]I'm not defining it. I want the LLM to surprise me with something. So, each of the four models is going to get the same prompt, and we're going to see which one comes up with the most visually interesting and the most fun version of this classic arcade game. And then the top two are going to move on to the final round. Now, we're going to start with Gemma 4. Gemma 4 did not do well in the last LLM battle. Here's a quick highlight reel. Okay. [laughter] I don't know what's going on here. It's having a freak out trying to read a file and failing. [laughter] Okay. [laughter] It actually died. It died on the server. Bad news is we're getting three three tokens per second.

[00:03:15]But I think I know what went wrong.

[00:03:16]Gemma 42B in particular, I was using a Q4 version of Gemma 4. We talked about quantization a few episodes ago.

[00:03:23]Basically, we compress the model. We give up precision in order to save uh disk space and VRAM. Uh but the smaller models from what I understand are more susceptible to that crushing effect.

[00:03:33]When you can when you compress it too much, you can damage the model. and uh running Gemma 412B at Q4. Uh even when you're just chatting with it, I noticed that it makes little spelling mistakes here and there, which is which is odd. I think it's it's being crushed down too much. So, uh to fix that, we're going to run a Q8 version of this that same model.

[00:03:53]The command we're going to use to start it, I've had a few comments from in previous episodes complaining that I don't uh specifically list the exact model and the exact setup that I'm using. So, here it is. Here's the exact command I'm going to use to start up Gemma 4. Uh each model today is going to be limited to a context window of 128,000. Uh these uh tuning parameters I got from Hogging Face for specifically for Gemma 4 12B. I am going to use MTP.

[00:04:15]We'll discuss MTP in a future episode.

[00:04:17]Basically, it's just a way to speed it up a little bit. And because this is a 14 gig model, I think Gemma 4 uh at this Q8 is about 14 gigs. So uh I can fit it easily onto the one card. So I'm not going to split it across both GPUs. I'm just going to start it up with uh with CUDA visible devices equals 1 so that it uses that uh Pro 4000 card and we don't have the overhead of splitting it across two GPUs. So, I'm going to start that up. Uh you do get an error when starting up this Gemma model. I'm told this is normal. In fact, it even says in the error text, uh this warning is normal during memory fitting. But it it is it shows up in red and it's kind of scary, but uh it doesn't seem to stop anything from working. So, Gemma 4 is up and running. Gemma 4 12B is up and running.

[00:04:58]So, uh, let's go to open code and give Gemma 4 our prompt and see what it can do. I'm deliberately starting with Gemma because I have the lowest expectations of this model because it did so poorly last time. Not a single version of Gemma was able to pass the uh the prompt that we gave it last time. And that prompt, I think, was simpler than this one. So, uh, let's see what Gemma 4 can do. I [music] boot it up on my own machine.

[00:05:27]>> Okay, that was fast. 1 minute 54 seconds. That was surprisingly fast.

[00:05:32]That was suspiciously fast. Let's uh see if this even runs. Oh, it didn't even write it out. It didn't even write it out to disk. What? What? Okay, that's that's that's weird. I shouldn't have to tell it to write it out to disk. I think didn't the prompt specifically mention that? It did. Output a single Python source file. So, that's weird. It just spit out the the code right here in open code and it didn't bother to write at the desk. All right, I will point out if anybody's keeping track of speed, we are getting 65 tokens per second out of Gemma 412B running entirely on the RTX Pro 4000. That's that's pretty good uh speedwise. It's not blazing fast, but remember this is a dense model. All 12 billion parameters are firing to uh to execute this uh this prompt. Uh wrote it out. Let's fire it up.

[00:06:22]Okay. Oh, look at that.

[00:06:27]Okay. Okay. We got basic particle effects. We got little thruster firing effect. It's It's pointing the wrong way.

[00:06:37]The thruster triangle is pointing the wrong way. [laughter] It's It should be pointing in the opposite Oops. direction of travel. Also, the ship is supposed to be an elongated triangle. This is a regular triangle, which makes it really hard to see which direction you're pointing, but it does work.

[00:06:55]Uh, let's try the bonus feature. What did we get for a bonus feature here?

[00:06:59]Hyper fire.

[00:07:01]Oh, you can have unlimited number of projectiles at a time.

[00:07:05]Oh, that's okay. Yeah. Okay. Oops, I shot myself. Okay, so that's pretty good. That's pretty That's surprisingly good considering that Gemma 412B couldn't even finish last time. So, uh, it did something.

[00:07:19]Uh, it's I wouldn't say this is great, but I wouldn't say it's terrible. I mean, it works. It It is, man. Friendly fire. Okay. Okay. You know what? Not bad. Not bad. Not bad. Surprisingly good for uh Gemma 4 12B.

[00:07:37]Okay. Uh, I've moved that one out of the way. Uh, all of these will be available on my GitHub in the AI playground. You can download them and play around with them if you want. Let's move on to a model called Quithos. I believe this one was post-trained from Quen 3.5 allegedly with uh a bunch of uh reasoning chains from uh Fable Claude Fable uh built into it. So, it's allegedly really really smart model, but uh maybe that's just hype. I don't know. Here's the exact parameters I'm going to use to start Quos. I'm using a uh MTP Q8 version of this model. And here's the exact parameters I'm going to use to start it up. Again, we're going to use a context window size of 128K. And because this is also a very small model, uh I'm going to do the same thing. I'm going to load it entirely on the RTX Pro 4000 so we don't suffer the overhead of uh splitting the model across uh two cards. Okay. And then same exact thing. We're going to go to open code, uh paste our giant prompt, and let it go to town. Uh this is a 9B dense model. It's a small model. 9 billion parameters is uh it feels like this is going to be punching above its weight if it can actually do this. Uh, but this is a fairly simple assignment.

[00:08:46]So, I I think I don't know. Let's see what we get.

[00:08:51]And if you're keeping track of speed, we're getting between 60 and 80 tokens per second, which is uh pretty good.

[00:09:00]Okay, Quithos appears to have completed uh and it did write it out to disk. So, let's uh try this.

[00:09:11]And h interesting. Uh I can't see my ship. I can shoot, but I can't move. And it crashes. Uh so, several problems. Uh let's uh let's go back to Quittos.

[00:09:29]Okay. So, I'm just uh giving a problem report to Quitto. We'll give it a chance to fix itself. Maybe we should limit it.

[00:09:34]Maybe we should have a maximum like three attempts. Three strikes and you're out. So this is attempt number two for quithos. Gemma got it in one shot.

[00:09:42]Although I did have to remind it to write the the result out to disc. Does that count? I don't know. I'm just making this up as I go along. There are no rules here.

[00:09:52][snorts] Okay. Okay. Uh Quithos claims that the problems have been fixed, but if you look closely, rotation thrust required. key down. Q up. That's supposed to be key up. It's making spelling mistakes. Even in English, I'm not super confident this is going to work.

[00:10:12]Here's uh attempt two from Quithos. And okay, I can Nope. [laughter] Nope. Nope.

[00:10:26]Okay, we'll give uh Quittos another chance here. This is what attempt number three. This is it. This is it. If you can't get it in attempt number three, then you're out. That's That's a rule I'm initiating right now.

[00:10:40]Okay. Ship should be visible and control should work. Let's uh let's just see what we get. Uh nope.

[00:10:49]Nope. Nope. [laughter] Nope. [gasps] All right. Quits is out. I can't believe Gemma 4 beat Quitos. Maybe it's just hype for uh for Quitos. Is this assignment too hard? Did I pick a I don't think this is a hard one. Let's uh let's move on. Uh okay, I've got a model called Ornith. Uh Ornith, I believe, was also uh trained on top of Quen uh 3.5, if I'm not mistaken. Somebody correct me if I'm wrong. Orth uh comes in a bunch of different sizes. Uh we're going to try the 35B. I believe this is a mixture of experts model, although I can't find the uh active parameter count on hugging face, so I I don't really know much about this model. I'm going to start out with these settings. We're going to run uh Ornith at Q5 and I'm going to allow it to split across both cards because this is a a larger model. Uh I I I think this one is 25 gigs. So, uh let's just start it up. Uh okay, it's loaded. It's using just under 10 gigs on card one and just under 18 gigs on card 2. 28 gigs total. Uh same context window size, 128,000. Uh so, let's one more time go back to open code and give it the assignment and let's see what ornith can do.

[00:11:58]And if you're keeping track of speed, we're getting 110 tokens per second out of Ornith. This uh it's because it's a mixture of experts model. It's uh it's I I I don't know what the active parameter count is, so I don't know how many parameters it's actually using at any given time, but uh 110 tokens per second is uh is pretty good.

[00:12:17]So, as I'm watching this work, I see that it's actually testing it by running the game for a certain number of frames.

[00:12:24]Uh, Quithos tried to do that as well, but it it kind of failed at it. But this thing looks like it's testing its own work by actually simulating user input.

[00:12:31]Uh, that's that's pretty good.

[00:12:36]Okay. Uh, Ornith reports it's complete.

[00:12:40]Uh, the bonus feature that it went with is no friendly fire. That's kind of boring, but okay. The disables friendly fire mechanic entirely, so you can shoot your own ship and it won't kill you. So, okay, let's uh see what we got. Here [snorts] we go.

[00:12:58]Okay. Oh, I can't seem to shoot and I can't seem to wait. What?

[00:13:12]Uh, okay. So, yeah. Okay. Let me I don't know. What is that? a projectile.

[00:13:23]Oh, it is. That's weird.

[00:13:27]If you hold space, the projectile just randomly appears. I don't know what the heck, [laughter] you can control the projectile. Okay, that's messed up.

[00:13:42]That's messed up. Okay, so that was attempt one. That actually didn't look too terrible. It It shaded the uh it filled in the asteroids. They have an irregular shape to them, but uh the controls are a bit wonky there. Let me go back and try to fix this.

[00:13:57]Okay, just going to give uh Ornith a quick bug report on that. And this is take two for Ornith. Uh every model gets three chances. This is chance number two. Let's see what Ornith does.

[00:14:11]Uh okay, it claims to have fixed it.

[00:14:13]Let's uh give it a shot. Just going to drag this over here.

[00:14:18]Oh, yeah. Okay. projectiles.

[00:14:21]Whoa.

[00:14:23]Oh, it Okay, it fires them too rapidly. So, you're allowed to fire three projectiles at a time, but even if you tap space, it's shooting all three at once.

[00:14:37]So, it's like a machine gun. You can't uh you can't shoot just one projectile. So, it does work. The asteroids are moving around and they Oh, that did that one just pass right through me. Let's test collision detection here. Let's see if I can hit one. Yeah. Okay, so collision detection works, but the projectiles are uh coming out way too. And again, we've got a perfectly regular triangle as the ship, which makes it hard to tell which way you're facing. It's supposed to be, you know, a lengthened triangle, an elongated triangle.

[00:15:14]So, let's give Ornith one more shot at that.

[00:15:19]Okay, so uh two bug reports. The projectiles are firing too rapidly and the ship is hard to uh it should be elongated. It should be obvious which direction it's facing. Let's see if Ornith can uh can fix this.

[00:15:36]Okay. Okay, this is third and final test with Ornith. It claims to have fixed the bugs. Let's uh see what we got. Okay, now the ship has a definite direction to it. That's good.

[00:15:49]Okay, much better. Much better.

[00:15:53]Much better.

[00:15:55]So, yeah. All right. Good. Not bad.

[00:16:01]Let's uh activate the bonus feature.

[00:16:03]There's no friendly fire, so I should be able to There you go. My bullets now pass directly through my own ship. That's That's good. That's kind of boring. It's not much of a bonus feature, but uh okay, it's it's acceptable.

[00:16:21]All right. I got to say it took three attempts, but uh Ornith's not bad. Okay.

[00:16:28]Okay. Uh let's bring in the reigning champion, Quen 3.627B, and let's try again. Okay. Uh Quen 3.627B.

[00:16:38]Uh I'm running the uh Q6. Uh this is from Unsloth. Uh this is an MTP uh capable uh version. So we're going to run with MTP. Uh I'm using these uh tuning parameters that I got from uh from Hugging Face and again a 128K context window. So let's fire up Quen.

[00:16:58]Uh this model is 23 GB. Uh okay. So, let's go back to open code and one more time, we're going to give it the same exact prompt and see what Quen 27B can do with this.

[00:17:11]And if you're keeping track of speed, we're getting 35 tokens per second out of this model. This is a dense model, so all 27 billion parameters have to fire.

[00:17:20]MTP makes it much faster uh by about 50% in my experience. So, I uh so it's going reasonably well. This is much slower.

[00:17:28]This is the slowest model we've looked at, but uh let's let's see what it can do.

[00:17:34][music] >> Okay. Uh so I just saw [snorts] Quen also did a quick headless test. It actually started the game uh programmatically and uh ran a few frames of it just to make sure it worked. So Ornith did the same thing. They both tested it. I think Ornith was more more thorough than Quen, but uh they both uh tested it automatically. I kind of like that. Oh, let's uh let's check out what Quen came up with.

[00:18:06]Oh, look. They got little fire effects for the thrusters. That's That's really cool.

[00:18:12]Oh, I can't shoot though. I can't shoot.

[00:18:15]Okay, I got to crash. Let me take that back. Okay, so quick bug report to uh Quen. This is Quen's second kick at the cat. Every model gets three chances.

[00:18:24]Come on, Quen. You can do it.

[00:18:30]Okay, it claims it fixed that. Uh, obviously it didn't include that in its test the first time around, but uh, it did the second time. So, let's let's try I love that thruster effect. That's so cool. Okay, let's try shooting. Okay, I can shoot. And yeah, we get Oh, let's that looks really nice. Okay, I didn't see what bonus feature it uh gave me.

[00:18:51]Let's uh toggle.

[00:18:55]Oh man, you can shoot. Oh wow. Okay, this wins. This hands down. Hands down.

[00:19:00]This wins. Look at this. You can shoot three bullets at once in try fire mode.

[00:19:07]Oh, this is this is actually pretty good. This is actually pretty good. This wins. I mean, holy cow. Look at that.

[00:19:15]Oh, that's I love that trif fire mode.

[00:19:18]This There's no way this isn't winning.

[00:19:19]That [laughter] Come on. Nice. Nice. Okay, so I'm going to give it to Quinn. Obviously, that was amazing. And I'm gonna give second place to Ornith, I think. I mean, Gemma did a a pretty decent attempt, but I'm going to give second place to Ornith because it did a really thorough testing job on his first pass. And Quithos is just out.

[00:19:43]That's just hype, I guess. Uh so, let's move on to the final round with our two finalists. Uh Quen versus Ornith. Let's go. Okay. Uh before we move on to the uh final event, uh something I should have done last time, which I didn't is uh I'm going to Was that something I said? I'm [snorts] going to get uh a code review of these three uh implementations. I say three because I accidentally deleted the one that was made by Quithos. No big loss because it didn't work anyway. But uh I'm going to get my code review agent to take a look at the three implementations that we have. the one from uh Gemma 4, the one from Quinn, and the one from Ornith, and just rank them in terms of code cleanliness and uh if there's any bugs in them. So, let's let uh let's let the uh code review agent uh take a look at these. Now, conflict of interest, the uh LLM that I'm using for this is Quen 3.627B. So, if it recognizes that it itself wrote one of these implementations, maybe there'll be some impartiality here going on. I I don't know. Uh let's just let it run and see what we get.

[00:20:49]>> [laughter] >> Okay, I didn't see that coming. This This is a lot of text here. Let's uh I'll I'll post this on the uh uh on my GitHub uh when we're done here, but uh the the short version is that uh so Quen 3.627B thinks that Quen did the best job. Maybe that's uh a little bit of a conflict of interest there, but okay. Uh surprisingly, it gives second place to Gemma and then Ornith, the bug magnet.

[00:21:17]I I didn't look at the code to be fair.

[00:21:19]I just ran them. Ornith needs a full rewrite of his math. Interesting. I wonder if that should change my rankings then based on the code quality. Quen thinks Gemma did a little bit better than Ornith. Interesting. Interesting.

[00:21:30]Okay, I will post this on GitHub. And now let's move on to the grand finale.

[00:21:39]Uh, okay. Executive decision. I've decided based on the results of the code review that uh our top three finalists are going to advance to the finale. Why not? I'm just making this up as I go along. So, I've switched back to Gemma and our grand finale prompt is to make a full screen audio visualizer. This is similar to what we did in the last episode, except this visualizer is actually going to respond to the music.

[00:22:01]It's not just an aquarium view. This is actually going to show uh effects, visual effects based on what the music is doing. We're going to do this in Python instead of Java. And uh I've set up this prompt. So, uh we're going to accept uh an argument that is the image to display. just a static image is going to be on the screen and uh an audio file to uh play uh scale the image to fit the screen start playing immediately. You can we're going to have to do some fancy stuff with the audio loading. We're going to have to uh I'm suggesting a couple of different approaches here. Uh based on my own uh experience with this, I'm going to let the L1 decide the the way it wants to do it. It may have sync problems if it starts a separate thread and ffmpeg to handle uh you know getting the PCM data for the uh equalizer bars.

[00:22:45]Uh but I'm going to leave it up to the LLM. Uh we're going to describe the graphic equalizer in somewhat vague terms. I want something visually interesting, but I don't want to tell it exactly what I want. I want the LLM to come up with something for me. And as before, we're going to have a bonus feature. If you press B for the bonus feature, it's going to be something.

[00:23:01]I've given a couple of ideas here like particle effects or image glitching or maybe gent gently falling snow.

[00:23:07]whatever. Just put something on top of it as a as a bonus and as a neat bonus.

[00:23:11]I leave it vague to see what the LLM can come up with because I'm just kind of curious. No further ado, I'm going to give this to Gemma 4 and uh we'll just see what Gemma 4 can make of it. Whoa.

[00:23:23]Nothing. Okay, that's strike one. Let's fix it. Fix it. Fix it. Strike two. This is it. Gemma, you got to get your act together. All right. Uh Gemma is out.

[00:23:35]You failed. No, we're going to reset and move on to Ornith. Let's give it a shot and see what we get. We get nothing.

[00:23:46]Wow, we get literally nothing. Okay. Uh, let me That That's weird. This is attempt two. Wow. Okay, it's completely changing the audio architecture. It went with option one of the two options I gave it, and it's getting frustrated that that's not working. So, it's switching now. It's rewriting basically everything it did before using option two. [laughter] I mean I mean as long as it gets it.

[00:24:10]We're not paying by the token here, right? This is all local. So as long as it gets it, who cares? But this is not impressive. It's just outputting the same thing over and over. I think it's stuck. I think it's stuck. I don't think it's coming back. Wow. Okay, Ornith is out. We are not doing well in the finale. This is not good. All right, let's move on to the champ. And Quinn is ready. Ladies and gentlemen, we are down to one competitor. Quinn basically wins by default if it can do anything anything with this prompt. Let's take the prompt and give it to Quen.

[00:24:45]Okay, let's uh let's see what we got here.

[00:24:50]Nothing is what we got. Was this too hard? Let's give [snorts] it another shot. Strike two. Attempt number three for Quinn. Wrong monitor and it's not animating or playing audio. Wow.

[00:25:05]Wow. What a disappointing end. Should I give it one more chance? I'm kind of cheating, but I really want to see at least one of these work. And Quinn seems like it's got the closest one. Maybe I give it one more shot.

[00:25:16]It No. No, Quen. No. All three contenders knocked out in the final round. [laughter] Oh my god.

[00:25:27]Okay, it's super late, but I don't want to end it like that. All three contenders were knocked out, and I think the problem was my prompt. I was uh I was confusing in my wording of that prompt. You can go to the GitHub and go to the commit history of that of this file. And you'll see the prompt that I gave them. And I think I confused things by giving them the two options for audio decoding. And uh neither one was really well phrased. So with Quen's help, I've rewarded the prompt. It now looks like this. I'm I'm I'm I don't want to spell it out. I don't want to tell them exactly how to do it, but I don't want them to get confused as to uh what is the correct solution here. So, I'm being a little bit more explicit in how this should be done. I'm going to reset and hopefully at least one of them will get it with the updated prompt. Let's give it a shot. So, we're going to start with Gemma.

[00:26:17]Okay, Gemma 4 claims to be ready. Okay, let's see what we get. Uh we get an error is what we get. This is getting deja vu.

[00:26:30]Okay, Jamma claims it's corrected. Each model gets three attempts at this. No more. That was attempt number two. And we're still getting an error on startup.

[00:26:41]The third and final attempt for Gemma.

[00:26:44]This is it, Gemma. Last chance.

[00:26:48][music] Okay, we got something. We got something. That's the best result I've seen so far tonight. Uh, all right. All right. Well, that was third and final attempt, so I'm not going to try to fix what was the bonus feature supposed to be, by the way. Did it say bonus effect? A base reactive color pulse. Well, whatever that was supposed to be, it didn't happen. So, all right, let's uh let's move on to Ornith. Okay, Ornith is loaded and ready to go. We're going to give it the same exact prompt.

[00:27:21]And here's attempt number one from Ornith. Come on, Ornith. Give me something here.

[00:27:28]Okay, Ornith is finished. That actually didn't take very long. That was I should be timing this. The last run, man, I don't want to edit this episode together. The last run, Ornith ran for 10 minutes. I swear to God, 10 minutes and then it got stuck in a doom loop where I couldn't proceed any further and it just died.

[00:27:43]All right, let's see what Ornith came up with. Uh, bonus feature is an auto audio reactive particle system. Particles spawn from the bar area. Shoot upward with gravity. Okay, that sounds like fun. Let's give it a shot and see if it even works. And nope, immediately we get an error. Nobody can get this one on the first attempt.

[00:28:01]Let's uh give that bug report to Ornith.

[00:28:04]And this is attempt number two. Everyone gets three shots at it.

[00:28:10]Okay, it found and fixed the bug. Let's uh see what we got.

[00:28:17]Oh, still not working. Ornith, Ornith, this is your last chance.

[00:28:27]That was immediate. That was very, very fast.

[00:28:30]But I don't know, man. Let's see.

[00:28:34]Wrong monitor. It's not playing audio and it's not animating.

[00:28:38]Terrible.

[00:28:40]Ornith is out. Gemma 4 has beaten ornith. Did not see that coming. Let's reset and try Quinn. Okay, new session.

[00:28:49]Quinn is loaded.

[00:28:53]Let's give it a shot. Go Quinn, go. You are the reigning champion. You cannot let Gemma beat you. That would be embarrassing. That would be embarrassing if Gemma 4 beats Quen. I'm going to be stunned.

[00:29:07]By the way, can I just point out that I'm really glad I'm not paying by the token here. This is all local. If I were paying by the token, this would have cost this this one evening would have cost me a a fortune, but because it's local, it cost me nothing.

[00:29:26]Okay, [cough and clears throat] Quen is finishing up here. The bonus feature is audio reactive particles.

[00:29:33]Looks like everyone's going with that feature tonight. Okay, let's give it a shot. This is attempt number one from Quen.

[00:29:42]Wow. Core dump. Unbelievable.

[00:29:45]Come on, Quen.

[00:29:49]Okay. Okay. [clears throat] Attempt number two.

[00:29:55]Wrong monitor. No animation. Okay.

[00:30:00]This is Quinn's chance to fix it. Let's see if we can do it.

[00:30:08]Okay, Quen found and fixed three bugs.

[00:30:11]Let's see what we get.

[00:30:14][music] It's a nice touch. You know what? I'm going to go I'm going to go just one step further. This is cheating a little bit because it's had three kicks at the cat, but I'm just going to see if it can fix those last two issues. the uh the equalizer bars are just a little too short and uh the bonus feature doesn't seem to work. I'm just going to see if we can tweak it. I'm cheating a little bit. Let's just go for it.

[00:30:41]Okay. Okay. Attempt number four for Quen just to see if we can get that bonus feature to activate and make the bars just a little bit taller.

[00:30:51]Okay. Okay. What do we got?

[00:30:55][music] Still no bonus feature. [music] That's really no better than what Gemma came up with. I'm actually surprised.

[00:31:07]I'm actually surprised. I got to give it a two-way tie between Gemma 4 12B and Quinn 3.627B.

[00:31:14]And Quinn took more attempts.

[00:31:16]Wow. Wow. So, Quithos was out in the qualifying round and uh Ornith was out in the final round. It comes down to Gemma 4 and Quinn 3.6. Both of them got the basic visualizer up. Neither of them got their bonus feature to work correctly. Two-way tie. Who so? Who would have thunk it? Who would have thunk it? Okay. Well, it's super late and I really really do not look forward to editing this mess together, but it is what it is. Thank you for watching that.

[00:31:45]This episode might be a little long, but uh it is what it is. We will see you next time.

[00:31:52]Will Quinn 3.627 627B defend its reigning champion title. Let's find out.

[00:31:58]That sounded angry. The two will advance to the final round where they will be given a more difficult prompt to