This Open-Source AI Taking Over 3D & Blender — Kimi K3 transcript

Who would have guessed that open source will take over that soon? Just a few weeks ago, Fable 5 was restricted, considering as two powerful. But today, open source Chinese model lands right into the top together with Fable 5 and GPD Soul 5 .6. How crazy is that? Not only open source, but also most cost efficient between all top models. And it gets so popular that provided needs time to increase capacity to fulfill the demand.

But there's one more claim that Kimi K3 is actually very good in 3D and developing 3D games. And I absolutely cannot ignore that. So in this video, I'm going to test how good it is really in 3D. We're going to test it in blender and game engines. I'm going to provide it with APR and all that stuff. And of course, we're going to compare results with Fable 5 and GPD Soul 5 .6. And of course, we're going to create complete game and see how good it is.

Let's see if hype is real. Now first, we're going to need Kimi to install. Of course, you can use Kimi as a chat GPT on the website or install Kimi desktop app. But that's not a way. I'm going to use harness instead and I prefer CLI. There's also a way to install it in VS code or I guess in cursor as well. But I'm going to just do this CLI. Since I'm going to work with clothing, codecs in one space. So that definitely works best for my setup.

I'm going to test it within my agent. There are 10 test prepared and 3 CLI opens within one folder, Clot, Kimi and Codecs. You can see all the settings that I use as promised. There will be Clot Fable, Max Thinking, Codex, Soul, Max Thinking, Kimi, Max Thinking, all that stuff. I'm not going to use plan mode until we're going to make the complete game. Test look like that. We're going to go one by one. Each basically has prompt, the image, or sometimes GLD model.

Also, all agents have access to blender mcp, full AI API with key and go dot mcp. In this video, we're going to do test in good dot. I think that suits better for one prompt test because to get nice results in the religion, it's a little bit more difficult and there are blueprints and stuff. But if you're particularly interested in Unreal Engine, check out my other videos where I show right approach how to use it with new Unreal Engine mcp.

I did a lot of research about that. And the first test to warm up will be to create a donut. Of course, that's the reference for amazing blender guru. Hopefully one day I will do collab with him as well. The test is simple. We're going to provide an image and ask it to do a render similar to that in if AI wants. It can go in watch tutorials. And part of the task was to do that within blender mcp so you can see pretty much how it has it happen.

I'm not going to show all processes how it does the job, but basically it runs the script and it's not really much happening. It's not that it goes through iteration, no, it just writes one big script and then do it. Sometimes it do iterate on the final results with the renders. Okay, now let's review the result. Here is donut from Claude. And this one is actually looking very nice. I think it looks very nice. Let's see what it does in blender.

Here is the blender scene. You can see it created each spring colesce separate object. Okay, so here we go. All right. Here it out looks like. Well, actually it's not that bad, right? A lot of lights, camera, etc. Let's see the keymerism. Okay, here it goes. I think it's pretty close, but in terms of colors, maybe a little bit too bright. Or is it designed that way? Let's take a look at the blender. Okay, so here we can see another approach to the background.

Actually, this is quite good. And if we enable that, so what we see here? Well, I think in blender, texture looks a little bit better. So I wish it can do a little bit better rendering. Otherwise fine, and here is the codex. Okay, so, well, I would expect a little bit better quality. I don't like the spring goals and again, like this look like completely treaty -ish. Right? So, doesn't look good. Let's see what's inside the blender that it prepared.

Okay, and what are these pieces here and here? I don't get it, like. I mean, I get the point, it's like a drip, but yeah, it doesn't look natural. I mean, at all. Well, in my opinion, Codex sort of failed this test, and Codex is a little bit better than Kimmy, otherwise, there are pretty much on the same level. Which is actually impressive. Let's keep going. Test number two is this. We're going to give them our shape of diamonds and ask it to create a diamond shader.

Well, this is actually very difficult to task. Especially since I provide the mesh, it's more difficult because it must create a shader that fits in a geometry. Usually, these shaders are very tight into the surface. Let's see what we possibly can get. Okay, so here's the Claude results. You can see the shape that I provided. Well, that's definitely not a diamond, right? And the classic way. But yeah, okay, it's something.

Let's check a look on the Kimmy. Okay, so, it's pretty similar to what Claude did, but it's like more rainbowish and it's like a disco bar rather than a diamond, right? Anyway, let's take a look on Chad GPD. Okay, this one is a fancy disco bar. Well, but to be honest, it's probably the best from what we got. I mean, it reminds refraction and stuff. And look at the sprinkles around. Well, I mean, that's a good one. So in this case, obviously neither AI did the job, but here I will give my point to Chad GPD.

Let me know which one you prefer. And guys, kind of reminder, don't forget to like it this video and subscribe if you like my content. I really appreciate your support. Test number three, we're going to take the character. I actually created this character one of my other video and we did all the clothing. Bye, bye, bye. Now the test here is to turn it into the hologram. I mean, you know, this sci -fi thing capsule hologram effect.

I will provide a GLB, go to MCPs working. Let's see how they do. First, Claude Caud. And result is actually very nice. It is actually very nice. I mean, something like this actually can be used in the game. Look, okay, the podium is animated. This like shiny, corn very nice. I like this a lot. Let's take a look on the kimi. Okay, we do have very, very close to Claude Caud. Well, there are clothes. I think it's like a little bit different approach to the podium.

Okay, so it's like a podium and there's like light goes from it. I'm pretty much the same animations. I guess like the cone would add more to this. And then we have codics thing. We will. I guess, you know, the codics in this case is the guy that wants it to do an extra effort. But it failed without those extra details. I mean, this apparently does this thing on top is not working. But the amount of cool stuff it tried to introduce is appreciated.

But yeah, it's like a failed approach. You have a cone. We do have these particles like floating interesting. I should work that out. So for this one, I think Claude and Kini passed and Jipiti, black for effort, but try better. Test number four is procedurally recreates geometry. This one is a little bit difficult because it actually needs patio understanding of objects. And here I'm going to ask it to use something cool.

I found recently like a skill called Image to Trigiers. Here's the link you can try it out. Basically it helps to recreate procedurally any shape and make it sort of interactable. Let's take a look on the results. And since the result is a webpage, I'm going to open it all here as well. Here is our Claude version. Ooh, well actually is very good. Like, okay, we can open. We can click the buttons. It looks very nice. Okay, okay, I like that.

We cannot turn it on. Unfortunately, by that's fine. Let's take a look at Kini. And here we go, Kini. Actually, the very close, well, yes. Nice, all the buttons here are right. Okay, and again, I want to point like this part is done nicely and it closes. And when I was checking like how it does it, so it does some shots very fine like whether it works nice and didn't do the iteration to fix that. And here we go with the ChaggpT.

Was saying before that ChaggpT had best spatial understanding like visual. But this example is absolute failure because the shape is wrong. Right, the buttons maybe are a little bit fancy. Like there's some stuff on the buttons, but the shape is wrong. There is a missing hinge that actually do the closing, etc. So ChaggpT enough passing this one for me. Well, Claude and Kini pretty much the same thing. Let me know if you agree or disagree.

Test number five is about simulation and not just simulation, but again, we know there are capable of doing some scripts. So let's involve some vision. My idea is that we do have a glass and rock, rock and the break the glass and charts must fall naturally, but also form the word kimi. How about that? Let's see how they're going to do that. Okay, so let's start with Claude Codes. Okay, do we have glass here? There is a stone boom, brakes and here we go.

So, woo! Interesting. If you can see that it's sort of like at some point of time at the tachias, this kimi charts to the back of the glass and then they accurately fold out. I think the only problem here is that it's too perfect. I was asking like to sort of like look naturally like accidentally forming the text. Anyway, let's go to the kimi. Let's see if it's better. Okay, pretty much the same setup. Okay, boom, let's go.

Okay, I think it's pretty much the same as Claude Codes, but the charts are smaller. Well, charts here look more naturally, but this stone is still here and rolling. So this one is like not good. Like, don't like the stone behavior. Otherwise, it's very similar to the Claude Codes. Let's see the charge APT. Okay, so here we go. The setup, okay. Set up is clear. Let's play it out. Dang, I mean, this guy is surprising me.

Look at this. Why is Claude Codes the opposite side? Ha ha ha. Well, basically, nah, no, well, it's bad, right? It's bad. Everything is here is so unnatural. No, I think Chagipiti failed it like in many ways, right? Like all charts goes into the text. No, like I really expected it to be like other did. Anyway, next one is animation retarget. And I already did that test. Two months ago with Claude Opus 4 .8. It couldn't do that like in few prompts, right now we're going to test it with one prompt.

The task is simple. We do have animated mixed summer character. And we have more character that I rigged with accurate. The skeletons are quite different. There's a amount of bones and the structure is different. But of course, there are human knowledge skeletons. So it's possible to retarget it nicely, while it will require some effort. So all AI did the thing. Let's see the results. Let's start from Claude. Here we go.

And I press play in. It looks good. I do not spot any issues really. It's pretty good. The animation itself is a little bit hard, so the limbs can tilt a little bit. But you can see that follow exactly the reference, which is I think very nice. Even the fingers. I mean, the thing with the fingers like ghosts like this, but yeah, it's fine. Let's take a look at Kimmy. Okay, here we go. Well, this is the same thing. Actually, it's the same thing and also really good.

I think it's great job. Right. I'm looking on this tomb again. Like, and I think like it did a great job here. And here's Chagipiti and Chagipiti also did that in one prompt. And Chagipiti proves again that it is more creative guy. Well, I didn't ask about podiums and stuff, but it added this fancy podium. Well, someone give this guy a medal. Mulgun, test number seven. Again, rigging and animation, bots, way more complicated.

Now, I'm going to give solid mesh of a dragon and ask it to rig it and try to animate deflying animation. This is their difficult task for AI from my opinion. Okay, so all AI done working. And by the way, I want to point one interesting thing. And that overall in terms of time, Claude Côte was defastus. Kimmy, somewhere in the middle, and Côtex sometimes faster, but in average, the longest. Okay, let's review the results.

Here we go with the Claude. Okay, here's the dragon. We do see the rig. You see the bones. And let's play the animation. Ooh, okay. What's happened to him? Let me check. So you can see in edit mode the mesh is solid, but I guess some problems, some issue with the weights. All right, let's take a look on Kimmy. Whether we see the same problems or no. Okay, so pretty much the same thing in terms of rig. Okay, well, it's not perfect, of course.

You can see some parts are moving with the ears. Like his ears moving with the wings right here, there and there. And yeah, here the weights are missing. But it actually is way better that Claude Côte, right? Let's see the Chejipiti and compare. Okay, this guy, of course he will add some small clouds and whatever this circles underneath. But it failed with the weights. Oh, whoa, whoa, whoa, whoa. I guess like it did some really bad approaching in terms of the weights, distribution, look, that's terrible.

Okay, so I guess Kimmy is the best of them all within this test. Of course, it's not complete. Neither completed it to be useful, but I think Kimmy was the closest. Okay, one more test before we're gonna do complete game. Particles, but not just particles, but how about the bonfire in tune shader? Well, like a tune fire, right? So the prompt is on your screen, let's see the results. Here we go. Claude Côte. Okay, okay, okay, okay.

Well, I guess it's not that bad actually like I don't like the shape of the fire, but there is smoke and there are particles, et cetera. It was fine. Let's take a look and Kimmy. Okay, so here's the Kimmy. I like how the light place, well, sort of like the real point lights are spawned, but there's too much of particles and it's like it's too nice, right? Should be like a bit lighter. Okay, not bad, but too dark. And here's the codex, Chejipiti.

Okay, I like again like this guy is always trying to add something I don't ask. I don't know whether that's good or not, but you see what you see. There is smoke, there is particles, et cetera. I think it's fine. Which one do you like the most? I don't really like any of them to be honest. But for some indie games, this can be very helpful. Right, trying to ultimate test, let's build the game. The first one is Mario Kart clone.

Why not, right? So the task is simple, two levels, three cards, and we're going to give it full AI, API. So it has to generate 3D models, the environment objects, textures, et cetera. Right, let's spend some money on it. Here is very important. I recorded like what settings I use. And in this case, we're going to switch to max effort, planning modes, whatever the model has to offer. In CloudCoded, it's called Ultra Code, which is like planning, plus extra effort.

For kimi, it's just enabling plan mode while the thinking is on max. So for cortex, we're going to choose more reasoning and choose ultra and let them cook. Now, for anything, when every model finished, they all missed to make the bat file to run the game working. Like all of them. So I had to ask it like why it started opening. And there was something like with the commands, because they were doing that in headless mode.

And apparently it's not working. Maybe that's a windows issue, I don't know. Anyway, they all fixed the dots. Fabel five did it for one hour, 10 minutes. Let's see if it works. Here we have the intraminium. Okay, we can choose two locations. Let's do that. Okay, the gate. Oh my days. So I see the card and look at the wheels. Like the card is so bad. Oh my, I'm stuck. Eeeey. How do I get out? Let's try again. I guess if you stuck, you start forever.

Okay. You can see the textures around. Okay. Well, at least there is nice surrounding. I don't know, like why it added the borders. I think the Murray card supposed to be open world. Oh my. So yeah, this is like a real issue. If you talk, you stuck like for real. Eeeey. That's cerebral. Let's see the ice level. Oh, there's snowman's. Oh, so nice. And yeah, the gate. The gate's place. It's so badly. Okay, okay, okay. So I guess that is Claude Court.

This is absolutely not playable, really. If you stock, I mean, I guess maybe it does the mechanics. Okay, now let's take a look on Chagip T. Okay, I like the screen, right? So there's like something happening there. Right. Okay, let's start with the beach oven. I look, everything is without the textures. No way. And why there is a card there. Actually, you know, I cannot, oh, it's so bad. Actually, I cannot control it really.

I didn't ask it for that, like I mean. So I can like just move forward and backward. This is impressing right. It goes left. I pressing left. It goes right. There is no free movements like way out. It's so bad. Let's take a look on the ice level. Okay, this primitive, so I don't think it's generated. It missed the textures while the fence is tough. I guess a rural light split. No, this is so bad. This is so bad. I don't really like it.

Don't really like it. It's practically, you cannot do that like it. Actually, all the place, right? Ooh, glass hope. Let's see what we have from Kimmy. Hmm. Okay, okay, we actually have images, right? Do you like this better than Chagipiti? I don't know. Like this one is actually reminding me, Marriacarp. Anyway, start with ice. There we go. Okay, okay, it's actually not bad. Okay, and let me check. We do have free movements and yeah, we cannot go.

Okay, I like we go through this strange. Okay, but let's see if I actually can move around. Oh, yes, I can, but I will die. Oh, look, I can start out drifts. No, not bad. Okay, boom. But it looks way better. Okay, I can play. Yeah, it's, it's very difficult to play. Definitely need some tweaks, but and I'm actually losing, right? Let's try the beach level. Here we go. Okay, fish level. I think the, the, the extreme on the road is a little bit off.

And there's like a small pond, no, I think they're a little bit inside the level. Let me check. How do we check? Yeah, this little of lags within the moments. I'm the chore. I'm actually winning. Oh, let's see if the palms are actually placed at wrong. Hey, well, let me die. Oh, it's like instantly just dropping me back. Oh, I actually can hit that and charge a PT actually fail that. Okay, so here's the results from one prompt.

Well, in my opinion, compared to Fable 5 and GPT 5 .6, Kimi definitely did the better job. Well, none of the results is actually a game. It's quite impressive how far we are moving. And as a reminder, again, Kimi is open source in way less expensive than Fable 5. Actually, almost four times cheaper. And the results are even a little bit better. Not like crazy better. Gonna be honest. I did one more test only with Kimi because I only had limits left on Kimi.

Fable 5 is very expensive. That is test number 10, rebuilding seems city. And this idea is from one of my subscriber, Travis. So this one is for you. Again, we give fully eye go to odds, etc. It can generate 3D models, texture, all that stuff. And I left it cooking for whole night. I cannot see the timer was cooking, but based on the context usage, I guess, around 4, 5 hours. Let's play it out. Here we go. All right, so this is like a sim city.

Okay, nice. I liked location. I was referring to sim city 2000. Just also because I was playing that game when I was kid. Okay, let me see if I remember how to play that. Let's start with the road, right? Let's try to build a, oh, there are cars. There are cars. It generated nice cars and they're like, wow. No, I mean, like, I know I should not be surprised, but I just like, okay. And it like, this works actually, okay, it's not perfect.

But I can build that. Okay, let me spit out by it. Okay, gonna do something like this. I'm really, really remember how to do that. Anyway, now we do have residents. Okay, let's do the residents here, here. And that's sure we're gonna build somewhere here, here, here. And why is nothing happening, actually? Why is nothing happening? Oh, I think I need to do the plants, right? She'll build, oh, wow. I mean, guys, I'm excited because it's a nostalgia hit.

I mean, that's why I'm so excited. Wow, let's build out a building, so I guess, school. Ooh, okay, so we like sort of quality of life hospital. Okay, and police station, let's over here. Boom, nice. And I think I didn't build any commerce, right? Let's add a commerce, boom, boom, nice, nice power. We do have not enough power. Okay, then I need to build another plants. I guess, like somewhere here. Okay, now we do have, let's do a commerce here, more residents, boom, roads here, how about another school here?

Ooh, look, big buildings. Ah, when I was playing the original one, as a kid, I was always wondering, how to get this nice big buildings? Well, I guess, now I can build it. Oh, oh, oh, wow. Like two major build buildings. Okay, oh, yeah, this is like the biggest building. I swear I did nothing to this treaty models, such a tidy, I just said, like, just do the models. And what I'm surprised is all buildings they fit these squares.

Because I was thinking it will be able to do the thing, but whether it will be like the real squares or not, and they fit the squares, which is cool. And we can actually have like speed up, we are running out of the gas again. Electricity, let's do that in nothing, let's just, and we need roads, right? So they can't go, I like it. Oh, and I can do like that, right? Ooh, nice, absolutely nice, oh, oh. How cool is that?

I'm absolutely amazed, guys, this is super, super fun. I think this type of game is the easiest to do. All right, that was KimeK tree. I think it's absolutely great to have an open source as a top tier model, because this news about Chagipiti, Fable 5th, being restricted was sort of concerning not gonna lie. And I believe that's one of the reasons that KimeK tree gets so much attention right now. Let me know in the comments what you think, guys, that was Steph, like and subscribe, and see you in the next one.