Kimi K3 Is INSANE – Is THIS a Sol & Fable Competitor?

It's lagging. Oh, yes. It has them. [laughter] I get so excited when the bullet holes actually show up in the environment. >> Thousand miles of this shot. [laughter] >> So, there's this new model out. It's called Kimmy K3. And I want to take a peek. No, I'm kidding. So, today we're going to be testing the long awaited Kimmy K3. This is exciting for a large number of reasons. And speaking of large, one of them is the overall size. This is a 2.8 trillion parameter model. It is massive. It is by far the largest is this open weight open source model scale. This is the largest openw weight model that has ever been released and that is extremely exciting. Additionally to that they mention prominently and in the recent days I've seen a lot of comparisons and folks saying this does match some of the outputs that have been seen with Fable 5 and GPT 5.6 Soul. They even mention this right here front and center. Overall, Intelligent is ranking only second to Claude Fable 5 and GPT 5.6 Soul for the complete benchmark result. See our tech blog. Now, that's not currently out yet, but nonetheless, we're just going to be more focused probably on looking at some outputs today for this model. Now, this is crazy because look at the size. And they even do have a benchmark chart here, or I wouldn't call this a benchmark chart, but just a size chart of open-source Frontier models, as they call them. This is significantly larger than the previous largest and that's I think kind of exciting. So while this really was just released like I refreshed the page this popped up this is like 3 minutes old but we have a bit of information even without the tech blog and one of those things is mentioning the important question the full model weights of Kimmy K3 will be released in the coming days. Additional details on architecture training and evaluation will be published together with the K3 technical report. That is awesome. And yes, most folks are not going to be able to run this at home. It is the fact that this is being produced and then put as an openw weights model at this massive level of size and supposed intelligence that is very very exciting. So before we go further, do feel free to subscribe as I want that 100k plaque. And with that, let's just take a little bit of an additional look at some of the things contained within this quick start post. They talk about some of the architectural considerations right here. This is of course a mixture of experts model. This is just meaning that it has a large overall parameter size of 2.8 trillion, but the active parameters will be less than that off the top of what I see right here. I don't know the specific breakdown for that, but I'm sure by the end of this video, we'll have a bit more pertinent information there. And they talk about some of the things that this is good at, mainly coding. Now, additionally to that, something that excites me is a lot of what's listed here, basically things I really like to do tests on in these videos, including CAD workflows. That excites me because one of the more recent tests we've been running have been basically making little CAD models of engines that can fit remote control motors and then seeing how well they actually faithfully reproduce something that fits a real functional part. So, that's an exciting test that I'm looking forward to running. Additionally to that, they just talk more about some of the benchmarks. This has a 1 million context token window, which is rather long. And again, I don't have tests specifically designed to really stretch that context window, but it's good to see that it has the capacity to go that far. Also, vision input, which is a big one. It does have the capacity to see images, and that can really help things when it comes to troubleshooting certain issues. Often times I do like just giving the model screenshots of problems that we encounter because it's one fun to see them work around just visually seeing what's wrong and then trying to fix it. But two it can sometimes make things a bit simpler if you can't find these specific words to describe an issue that you're noticing. Now one of the big questions folks are going to have being that even if this is open weights it's not very realistic to run for most folks the price. So, this is definitely more expensive than a lot of folks are likely going to be expecting just based off of previous models from the open weight providers. However, if this intelligence truly does stack up, then this listed price here of $3 per million input tokens and $15 per million output tokens will still be a decent deal considering that 56 soul is, I believe, $30 per million out and then Fable 5 is 50 per million out. So, it is more expensive. However, if the intelligence does stack up to those models, it is definitely cheaper than the alternatives. So, this test may be a little less structured than some of our traditional ones because I'm going to be running multiple multiple queries in parallel. But, nonetheless, let's just start with the triedand-true browser OS test v2.5. Is this outdated at this point? Yeah, probably. So, I'll go through this a bit quicker than normal. Nonetheless though, it's always fun to see what models do, if there's anything that surprises us. And I am going to be using this for the specific test just through the web chat interface. K3 is selected and it is on max reasoning mode. I do believe I read somewhere in that quick start that it currently only supports thinking mode and max is what it defaults to. So that is all correct there. And I don't find that this is really going to be needing an agent swarm to produce the browser OS. So we'll see what we get. I'm not looking at it. Like I can see it's open, but I'm refusing to look at it. So you'll have seen it prior to me. All right. Very good. All right. So, here's our Kimmy K3 browser OS. Hi, I'm Pulse. Double click to pet me. Double click for cookies. Oh, click to pet me. Double click for cookies. I was like, that's logically not possible. Okay, so there is no right click. Nonetheless, let's let's double click this thing. Happiness plus one. Oh, good. No. Yum. Crunchy. He again. Again. I needed that. Cookies fuel my circuits. Om nom nom more please. Oh wow. Okay. That's quite something. So we have a clock with the correct date and time in our local again just in case you missed it. No right click. Nonetheless, let's check our Nova start menu. Okay, we have about synth pad. I like when I see stuff like that. I'm going to just run through them top down. Oh, how do we get pulse off of our screen when we play GTA? Okay, this looks incredibly impressive. Now, I said that with the 5.6 soul result, and I ate my words real fast. I just need to make sure the speaker's working and then we'll test it. All right. Do you think this thing makes noise when we All right. Start engine. You know what? It's It's not bad. Good. That looked expensive. Is this thing commenting on our GTA behavior? Oh my goodness, it is. Well, why would that have been expensive? We ran someone. I guess lawsuit wise, that looked expensive. This is very like 2.5D. I feel like the walking animations and the pedestrians are actually more impressive than we're able to notice. The driving is smooth. Okay, wanted level down. We're going 97 km, so it seems like the max speed of this car is a little under 60 mph, which is quite pathetic. Nonetheless, the police are chasing us. Now, something that differentiates a model that's like good versus great. If the police cars snap to our point and then just overlay our car, that shows like a lack of intricate thinking. So, we want to see if that Okay, good. When they hit us, we just get busted. The 5o got you. Ouch. Thanks, Pulse. [laughter] This like little cute thing is like talking like street. It's like, oh, the 5o got you. This thing was just like nom nom nom. when we gave it cookies and pets. Like, wow, that's quite the All right, this is this is cool. The buildings are drawn. Well, I don't think we can get out of the car. Run over pedestrians for cash. Dodge the cops. All right, good. Okay. Oh, no. A pedestrian dot. This thing's actually kind of entertaining. Now, I'm glad that we weren't able to banish it. Happiness. All right. And then the police cars do have sirens that show like blue and red. Can we minimize and reopen? Good. I'm happy with that. Grand Theft Auto Neon closed. I kept all its secrets. Thanks, Pulse. Voxilcraft. Oh, these are always the kind of demos I see people posting on X. Just like these total like zero creativity thought demos. I'm kidding, by the way. I'm being somewhat sarcastic. All right. This is actually kind of sick. Now, I'm going to probably someone will likely judge me for what I'm about to say, but I never played Minecraft. I never got into it. Yes, I've played it like maybe a total of 20 to 30 minutes my entire life, genuinely, but this seems pretty cool. We have different things here. Are these supposed to place them? WD to move, drag mouse to look, left mouse button, right mouse button to place. Nice. This thing's not commenting on what we're doing. F to fly, space to jump. All right, this is actually kind of cool. And the blocks change from the number pad. So, one to six. In terms of a second game in the browser OS, this is probably the best second game that's ever been included in the browser OS by a long shot. All right, this is cool. I'm spending more time on this, but it's the only result so far that's finished. All right, this was actually kind of awesome. Next up, terminal. Okay, we have wallpaper settings there. We'll save those. That's good looking. Nova OS Novos. That's okay. Pulse pet on off. Oh, perfect. That's messed up. I'll I'll bring it back. I like it. [laughter] I've started to warm up to it. I'm back. Did you miss me? Absolutely not. Clear. All right. Simple terminal, but nonetheless, let's check our paint. Can we resize without having the canvas get like messed up? Yes. And that is something that surprisingly is not always the case. I'm I really need to like start a like a different channel. Bye Paint notes. Welcome to Nova New. Okay, good. It has a nice sidebar and we even have note titles. That's something rarely if ever we see. Can we save this as a text file? Okay, I don't see the ability to do so. Nonetheless, still very good. All right, synth pad. Opening synth pad. I shall supervise. All right, kick. Snick. Can I loop this thing? [music] [music] >> [music] >> Not bad. I would have liked a way to loop the drum tracks, but not bad. Calculator. All right, check this out. Pulse 82* 6 492. Accent color. Do they work? Good. They do change the accents. Nothing so far has changed the pulse companion. Okay, we do have a toggle to turn that on and off. Good accent color. I do like the lime green. Let's take a look at our different backgrounds. Okay, live ones. Starfield, Matrix Rain, Wavelines. That's unique. And Aurora, that is also kind of cool, but also maybe a bit too fast. might not be like ADA compliant to have something flashing to that degree. Nonetheless, clock seconds. Okay, just toggles that on. And our special feature is, of course, pulse. So, next up, we're going to be using open code for a C++ skateboard game test. I am beginning this from within plan mode, and I've slightly modified this prompt, so it's not the same as any one we've sent before. This is the traditional self-contained C++ skate game. However, the environment is now supposed to be a late 90s mall map, walking NPCs storefront mall aesthetic, and a bunch of scenery to perform tricks on. I believe there was some Tony Hawk game that had like a mall map, and it seemed pretty cool. Additionally, I have just added to the end, don't overthink. This can potentially think for longer than is really like what we want for trying to test a bunch of stuff. So, with that, we're it's going to check on the system. That's what it's doing right here. just seeing, okay, what does this system have available that I can use to actually create this and compile it and build it? It's likely going to want to initially plan out to use Rayb. I would assess it will ask us some follow-up questions and then we'll be able to veto that decision. For those who are perhaps less familiar with some of our tests, the reason for doing that is not a disdain for Ray. Rather, it can simplify the model's ability to make something really good. We don't want that with big beefy models. We want to make it harder for them to see more raw capability. So, we'll allow this to think. It will probably come up with a plan and ask us some questions. All right. So, it opted not to ask us follow-up questions due to the fact that I specifically said don't overthink. So, that's on me, not on the model, just to make it clear. So, I've swapped it into build mode, just said don't use Ray. Everything else is good. Now, build it. All right. So, I can already tell you just from what we see right here, this is going to be pretty cool. It's currently drawing. I can read this. So, it's just drawing like the mall layout and stuff like I'm totally kidding. I've you know it's it's it's building this in a more hardcore way than it would have. So basically what it is doing right now is it's designing its own font I think to put it simply and we can see right here ABCD E F G H I J K 0123 to 9. So that's actually what it's doing good enough. Any character not in the table draws this question or skip font storage index by a string of supported characters. So this is what we mean when we say don't use raillay because we want it to be significantly more difficult things like this. All right. So basically just one of the infuriating things about this is multiple times now we've seen okay let me write the code and then it just goes back into thinking it's it overthinks big time and I'm going to make that determination just based off of this one thing that we've been seeing here so far. That's not behavior that's likely to turn on and off randomly depending on the prompt. It's more of like a core behavior. It's there or not. And in this case, at least for the time being, with this defaulting to the max reasoning effort, this is overthinking and it's just a little frustrating because it's going to end up taking probably like an hour to get a C++ skate game. It might be incredible, but you also do have to at some point factor in the time it takes to receive a result compared to like the quality of the result, I would say. Okay, enough planning. Write it one more. Now right actually also done zero warnings compiles clean one fi one one file runs stable all right let's check it out now let's play it for real good okay good mall aesthetic the pedestrians are walking now let's try to get on a rail here that's kind of cool and it does actually have logic there for whether or not we succeed okay that's Cool. The that trick was actually sick. Those effects are pretty neat. The way the board is spinning. This is arguably I would say don't think I'm glazing this. Yeah, the fable result like slapped this. It's not in the same comparison in the same league. At least with that though, this is a different prompt. This is the mall aesthetic one which we've not run before, but I'm going to say in terms of openw weight models, which this is, the board movements and effects are very smooth. Even the way that the player moves in tandem with the board as they would were this real life. I am actually kind of satisfied with that. So, let's see. I'm just trying to read the like pricks. Land on rail to grind. R to reset. S is to grab while in the air. That was like a split or something. I would imagine we'd need a bit of a higher. Now, the text is inverted. That happens a lot. It's kind of just something that we deal with when doing this. All right. Oh, the All right. So, we did a bail to grab. There's a bench. And we can grind on some of the like nonspecific rails in here. So, it's passable. It's It's close, but that guy got flexed on. See, that's okay. And the bales work well. So, all right. Overall, I'm happy. And we finally got a result. Mr. 89. I'll take it. In the meantime for that, I did from within the web chat interface here just have this do a front-end test. I should say from within open router. So, we're going to check that out. It is for the slapis watch website. So, here's our beautiful front-end watch website test. This needs to have a 3D model of a watch with a keyshot style render. I'm Okay, so this is actually really very impressive. The big problem that we're going to notice here is unfortunately the band is oriented around the wrong way. So it's it's like this far away and this was done through open router. So it didn't use perhaps some of its agentic capability that it would have were we using it through Kimmy's own front end or something like that. However, just looking at this watch face right here, this is really very very good looking. We have a secondhand moving the materials look nice. The wood on the table that it is to be placed on looks very good as well. Even look at the back. There's like engravings on the back case right here. We can kind of move it around. And when we scroll, it spins the view of the watch. And the rest of it is just the kind of traditional light white and gold. Obsessed with the last micron. Wow, that does check out considering how much this thing's been overthinking in the C++ gate game test. Nonetheless, the collection. Yeah, that looks very good. Okay, so it's 4850 and we see a nice just like scrolling effect on it. Really, the only issue here is that the band is oriented in the wrong orientation. Were this actually a properly laidout band, this would arguably be one of the finest results. Definitely decently in the scope of like GPT 55 or 56. Look at that. And then we have the green and gold one as well. This one's more expensive. These look so good. Drag to rotate. It is just frustrating about the band, but let's ignore that for now and focus on the good, which is everything else. The crown looks good. The emissive properties of the glass over the watch looks fantastic. It even has a little date thing there. It says Slapus on the face. There's a gradient to the green face right here. The second hand is moving. Okay, now that I look at it, the second hand may be a little off center, but still the model that it made here is very, very nice. That's interesting there as well. And we have a [clears throat] nice footer, Swissade, etc. Pretty cool. And this was just from open router. All right. And I'm also going to be giving this a 3D model test as that's something they mention it's good at. So I'm asking it for a 3D model of a V8 engine for my RC car project. It needs to fit a 280 size DC motor inside to add some detail to my build. It needs to fit nicely inside as I said and it needs to be detailed and something that is 3D printable without supports or weird geometry. So, we want to see if it actually properly designed something that can realistically be 3D printed without being a huge mess or requiring a ton of support. All right, your V8 engine is designed, verified, and exported. Everything lives here. Let's take a peek at this. I'm really very eager to see this as this is a test that differentiates the true models from the fakes. No, I'm kidding. But still. All right. So, here are all the parts separately. I'd like to see a render that looks very good from the side selection photo we see there. There's a 50% cutaway. And this is just now giving us photos of individual elements of the engine, including the fan. What was that? Uh, okay. I don't 100% know right now, but I'm sure we'll find out later. So far, this looks good. I think we need to see it in like um OpenCAD so we can get a better look at this. Yeah, that's that is pretty nice. And the thing is the pieces are not necessarily attached here in this render. And that is okay because they're designed separately. So, they're supposed to be printed one by one and then assembled together. Now, the big question is essentially the motor fitting in here. Is that going to work or not? And for that, we're going to need to look at the pieces individually just to make sure about that. Although, the preview may show us. So, here's the block. Yep, look at that. Son of a gun. This did something interesting. Now I don't know that that would have been necessary but if you recall I said make this in a way that really limits the amount of supports it needs. It basically in lie of doing the full circular shape here making this triangle made it less likely to need supports just in that gap. I wonder if it did that on purpose. If it did that is extremely clever and I don't really see any other reason why it would have done that and I'm quite pleased with this. Now, the thing is, this model is so slow that we can probably print some of these now because quite frankly, um, I've given up on getting this video out in any sort of timely manner. So, why not have some fun? So, I'm going to pop these into a flash drive and then slice them and then print them on some printers. I have a couple printers up and running, so we'll be able to get it done quicker than it would be if it were just on one printer. And we'll take a peek and see how they look in reality. But overall, it does seem like a pretty decent result on first glance. All right, so this engine model was absolutely stellar. Seriously, the pieces were wonderful. It had very few issues in terms of actually being able to be 3D printed without any supports. The motor fit inside the engine perfectly. There was actually ample space to run the wires from it as well. The cover to hold the motor in wasn't quite sized correctly, but it would have worked properly. And then all the accessory pieces like the fans, the air cleaner, the top of the intake manifold. This was a seriously seriously proper and impressive result. I am very, very, very impressed with this. All right, so we're now after quite a decent bit of time getting our Subway FPS result. Now, we're going to see a lot of stuff went on here just throughout the point where we sent this and then to the point where it's actually being delivered to us now. So, we're going to see it right here. I'm just going to open full screen preview. Okay, very interesting. That's a very GPT5.x looking UI. Everything preceding 56 looked kind of like this. We can either explore or we can play the zombie game. That's actually cool. Put some thought into that. Okay, this is that's excellent. That is quite quite good. The footsteps. This is fantastic. Look at the stain there with the wet floor sign. Night City. This is Lumen Station. There are mesh colliders on all of the things. It even blocked off the thing that would get us out of here just for some realism. This is fantastic. This is truly very very very good. Took long enough, but okay. Um, this is interesting map. No signal. Look down there. It's rare to actually get like the subway tunnel. A lot of times it'll just be totally blocked off. Staff only and then something in I would presume Chinese, but nonetheless, look at the lighting and the reflections. This is really really really nice. There's conduit. The ceiling is there. It does have security cameras. The exit sign. And when we sprint, the footsteps speed up. This is big, too. All right. And we can't go onto the tracks. Coffee news. Way out. All right. Let's try the uh switch mode. Look at that. That's That's just good. Okay. Oh, okay. Ah, it's lagging. Oh, yes. It has them. I get so excited when the bullet holes actually show up in the environment. Look at that model. And to reload, it took the clip out. Where are they? Oh, yeah. Here we go. This may be if not the best like second place. I'm thinking this might be the best result for this ever. Oh man. Look at that. Did we just pick up ammo? F is for flashlight. Oh, look at that. A medkit. So, we get drops as well. Wave cleared. [music] They smell your blood. Wave two. All right. All right. And they have Oh, wow. Okay, there's more of them. They're coming in like crawling. Hey, look, a fire extinguisher. Can we Oh, yeah, we can. This is good. This is what I was hoping for. My extreme rage from the skate game test has been more or less resolved. So, that's good. Oh, infected down. Wave cleared. I don't normally play them this long. And I think that should be somewhat indicative of how I truly feel about this result. All right. What happens if we lose? Let's just let it happen. Interesting. Look, half the arm is missing on this model, which was more like what you would expect from a zombie. [music] Overrun main menu, leave the station. That was excellent. Really, really quite well done. So, again, from within Open Router, I'm giving this the task to create an accurate 3D model of Jerry's apartment from the show Seinfeld. I have also included a photo that I have used consistently when giving models this test just as a floor plan reference etc. This will be an interesting one because historically there's been a rather large gap between open weights models and closed source proprietary models in this specific test though I have not done this in a little bit of time. So it'll be interesting to see how this handles it. Now just through open router I had also given it the test to recreate the apartment from Seinfeld. >> [laughter] >> All right. So, on first glance, as usual, it's not proper to the actual floor plan layout of this apartment. No model has been. So, we tend to just judge them independently based off of the individual assets. Instead of trying to judge this based off of now, it is possible. I mean, if this was like a hundred billion parameter and it had done this, I would say this is benchmaxed on this specific test. No if ends or buts about it. This is a nearly three trillion parameter model. So, it might genuinely just be this good. There are, yes, some issues here, but I mean, look at the way. Let's go into walk. Oh my, this is very impressive. I would say just the amount of like color and things like that. Sports Superman. We have a carpet there. The TV, the old style TV. This is likely the most lucid bicycle model I've seen in this test. What make of that what you will? Daily news. The table. The lighting is nice. The bookshelf with the little Superman character. [snorts] We can even shift to run. Well, I can't click doors. So, can I like move over here to All right. It's difficult because we can't physically make it into look at these other areas here. Okay. So, we'll look at it right here. No, we need to get into the bedroom. All right. So, we'll just zoom in like that. Look at that. Look at even the the differentiientation between the tire material and the rim material, the bathroom with the tile floors, the sink, the shower. Again, this does not look like the floor plan of the apartment. Basically, no model has ever done it to that degree. I don't think I've tested this with Fable. The chair looks really good. The bedroom, the vintage computer, the window, the paintings on the walls. Let's see if we can get into the bedroom. I don't think we can. That's frustrating. The bookshelf even like click doors. Okay, now we're in the bedroom. just kind of like inverted, but we'll we'll make the best of it. Look at that Porsche 911 Carrera. More or less the shape of a [laughter] maybe not, but you know, 742. Look at the clock, the bed, the Z fighting of some green material that was inevitably placed in the bed as well that we can't see. The outer windows, and can we open that door? No. Still, this is I mean that's that's really good. Fourth wall and ceiling. And then basically can't see anything outside there. And we can toggle that on and off. Very good. All right. So, next up, just from within build mode, I'm giving this another test that we have some decent and recent there comparisons to Fable 5 and 5.6 Soul. This is to create the 3D city block that shows us it in different time periods and the specific items and niche details must change to reflect the selected time period. So I have included as well 2055. It's always cool to see what they assess the future will look like. Pretty much every model except Fable 5 was fairly bullish about the year 2055. So we'll see what we get here. All right. We have our city timeline task here. Okay. Okay. Good. This is so so so similar to Fable or was that 56? Either 56 or Fable like massively so. Okay, it starts us who are just orbiting. That's a good thing. This is good. 5G everywhere. The future is now somewhat concerning. It put in music. Look at the people on their phones. And the walking animations are acceptable. A honk. The cars have LED strips for their lights which are more futuristic and modern. There is some curvature at least to the like windows. Can we click on things? Okay. No. All right. Those cars are morphing, but that may be a futuristic thing. Cold brew, escooters, 5G store, electronics, salon, roastery, roastery, bobati, gym, a fire escape. That's kind of cool. And the tops of the look at in between the buildings is like one of these. I don't know what you would call this. I'm sure there's a simple term for it. I can't think of it right now. It's like a normie space. I'm kidding. So, there's solar panels on the roofs as well. Pretty cool. All right. And just some more folks walking. So, I guess [music] let's start with 1945, which nice transition effect by War Bonds. Okay. Very period correct. Let's look at the cars. Definitely older style. The people have clothing on that is reminiscent of the time period. Right there you can see they're not holding phones. Interesting seeing the flags hanging around the street. The light poles have changed. The signs as well. There is still the normie space in between the buildings. I don't know that that would exist in 1945, but we'll let that slide. Grocery, barber shop, hardware, savings bank, radio repair, radio news, cola 5 cents, and different soundtrack as well. The buildings look good. The faces just with the some windows are lit up, some aren't. The cars are still morphing, but we'll let that slide. All right, let's check. Let me zoom out. Let's check 1965. I do like the transition effect. Oh, yes. Look at that pink Cadillac with big like um how am I I can't think of the term. Fins is the term I was looking for. And the cars are all very happy as are the colors. The people have the way they walk is interesting. Space age diner. Milkshake and burgers 25 cents. New color TV. TV sales. Everything's happy and bright and somewhat pastel. [music] Groovy cola. Things go better with it. Okay. The street lights are more like mid-century, I think. Let me just check that real quick. Okay. No, they stayed the same, but that's okay. The buildings got a little taller, I do believe, as well. I think they did. Yeah. Yeah, they did. Laundromat, florist, hardware. All right. 1985. Oh, yeah. That transition was kind of sick. Everything's neon. Video rentals. Be kind. Rewind. Look at the glowing sign. Cola classic. Catch the wave. Let's try to catch one of these cars. Okay, cool. It's got like a slope front. Very reminiscent of the C4 Corvette and other sloped vehicles from that time. Pizza. The people are Oh, this person is blue. And that's, you know, aerobics, donuts, arcade, arcade, and the music changes. All right. 2005 should have like bland corporateness. DVD World. Coffee to go. I do see people in baseball caps. That's actually pretty interesting there. The cars are a bit more modern and toned down. Cell phones, donuts, gym. On the roof here, we don't have any solar panels, but [music] the buildings do seem, at least some of them, to be a bit taller. Yep. Good. Is that a British phone booth? I dig it. I didn't really tell it where this needs to be, so I'm cool with that. And we have our horns honking back now. All right. It's very alive and active. 2025, which we did see because it loaded us into this. This was definitely more modern and nice. Now the big one. What does this model think 2055 will look like? Buildings are tall. Okay. Yep. Straight up like 85 retro futuristic style. Synth cafe. 100% lab grown flavor. Neuralink V9. Think it, stream it. Mars tours. 3-day weekends on Olympus on Mon. Someone can Nice. Nice. The normie space is significantly better just with the introduction of this cube holograph thing. Hover pods, holograms, and tickets to Mars. Let's look at some of the vehicles. Okay, so they're hover pods. Crypto Bank, Ramen X, Gan Style, and the people all seem to have like um Apple Vision Pro version 5s on them, but they do have like a bunch of LEDs on their faces and things. We have the hover pods. Is there anything flying in the sky? Yes, there is. So, it was hard to see there cuz it just went around the corner, but there were drones actually flying around. We see one right there if we try to zoom in on it and catch it. That's pretty cool. All right. I like it overall. And it's interesting how it changed the background as well. Interesting. The background didn't really get exciting until 2055 where it lit everything up. Cool. All right. Overall, this was very well done. And that took about an hour, just in case anyone was curious. All right. And I'm giving it this photo of this book cover and saying create an insane in all caps which will be in the title of this video as well. Front-end website for this book. The image is the cover called One Smiles. The site needs to have an insane high-end aesthetic. Quotes from the book. Things to hook readers. Pre-order signup. Author photo and bio, etc. Make it look amazing. Go all out. So, we'll see what we get for some more front end. All right. All right. Good. One smiles. Oh, wow. Okay. A novel by Nora Ellery. I think Nora is a name that's been like used a bit. Interesting. It's put almost like this visual snow moving effect over this. I don't know if that was done on purpose to do like immersion. Now, something else that's interesting is you see this cursor and the circle that's around it. If we go to the Moonshot website, that's almost like a one for one of what they have here on their homepage. So, I just find that kind of interesting that it's used that here as well. March 9th, 2027. Oh, okay. So, it's coming out later. 304 pages pre-order. One of them is smiling, one of them is carrying the rest. Oh, wow. Okay, we have the image again. Daniel, 26 years of marriage. One smiles. Interesting. It's zooming in as we scroll down. One carries the rest. Okay. Publisher Marian House comes out March 9th, 2027. 304 pages. We have an ISBN hardcover ebook and audio for readers of Anne Patchet, Elizabeth Stout, and Jonathan Franen. Oh, this font is rather difficult to read. Daniel Ren has never met a room he couldn't warm. 26 years of marriage and he still laughs first. Longest loudest still finds June's shoulder whenever a lens appears as if happiness were a reflex he could lend her. June Ren has spent the same 26 years quietly memorizing the exits. When the family returns to the house on Heron Lake one summer, long ago, never discussed, comes back with them. Their son is engaged. The portrait is commissioned and in the half second before the shutter closes, Daniel and June must each decide what face they will wear costs the other. And whether a marriage performed is still a marriage lived, told in two counterpointed voices, one smiles is a devastating, darkly funny excavation of a life shared. The performances we rehearse, the griefs we ration, and the terrifying hope that love might survive the telling of truth. Now, this background effect is horrible. This is genuinely exactly what it's like to have visual snow every day for your entire life. I have to say this isn't as good as I'd hoped. It's just Oh, okay. Very good. Now, this is excellent. The author page is very good. Here's Nora Ellery, Hudson Valley, New York. The author of two previous novels, The Cold Room Upstairs and What the Like Kept. Her short fiction has appeared in the Dunmore Review Harbor Like Quarterly and Elsewhere and won the Callaway Prize. She teaches narrative writing at a small college in the Hudson Valley where she lives with her family and a dog who has never once smiled for a photograph. Also by Nora Ellery. Okay, this is interesting. Be there when the shutter closes. Pre-order. And then we have a countdown timer. Barnes & Noble, Amazon, Water St. Okay, interesting. It's not as good as I'd hoped it to be, truthfully. Maybe it was garbage in, garbage out, but still, you know. So, all right. So, at the very beginning of even before I started recording this, I'd given this a prompt that I did in my Fable 5 versus 5.6 Soul test. I ran this same exact prompt using Fable 5 on Ultra Code and Soul 5.6 on Ultra Reasoning Effort. So, both of those were in very heavy duty, the highest potential reasoning modes. I've run it here from K3 within the swarm. So, this is the agent swarm mode. And I started this a little before 11:00 a.m. and it is now 3:00 a.m. Uh 3:00 p.m. I don't specifically know when this finished, but this took somewhere north of like 3 or 4 hours. Nonetheless, this is supposed to be a very, very high-end cinematic game for Steve the PC repairman. In giving it this prompt, I also gave it an OpenAI API key and said you can use that for voice assets and sound effects and things like this. So, it should have been able to create the speech assets that it needs for the characters. And basically, we're going to just go ahead and now play through this, keeping in mind, this is supposed to be a very, very high-end cinematic gameplay result. Okay, this is already very good. All right, so it should have cutscenes and like sound effects. Oh, yeah. Okay, so we start with an objective. That's Mrs. Ellis there. And it seems like she's um got her hands slightly detached from her body. We'll just let that go. Certified. Definitely not a fixer. Okay. There's a lot of like interesting sound effect here. Steve's PC repair. The cat. Steve is supposed to have a cat. So far, this actually in her car is right out there. This is good. This is better than the GPT56 soul result, honestly. Um, not better than Fable, but okay. So, we have a task to accomplish. I would imagine that. Read the certificates. Oh, Easter eggs. A+ certified. Close up the tower. Look at that. Admire the workbench. Okay. Oh, okay. Now, what? Pet Pickles. All right. I'm assuming Pickles is probably not the old lady, but the cat reminds us of our our old friend Pulse. Check her bowl. Okay. Oh, ring Miss Ellis up at the counter. No, I don't want to fax something. Someone faxed you a script page. It's about you, Meta. >> [laughter] >> Okay. So, sounds not quite right. And oh, it's not playing the sound. >> Stevie, you saved her. Harold's solitary scores are all on that old girl. >> Just a tired little battery. Males keeps the clock honest. She's good for another 20 years. >> That big store on Fifth wanted $300. 300 for a battery >> for you. 20. And don't you dare argue. [laughter] >> And a plate of my snicker doodles. >> Final like >> you need to eat something that isn't from a vending machine. >> The arms. >> Oh my. That's a serious looking fellow. That's just Mr. Thomas collects computers. Very private collector. Come on, I'll walk you to your car. [bell] >> You're a good boy, Stevie. [laughter] >> Drive safe, Miss Ellis. Cookies on Friday. >> Oh my gosh. >> I'm going to Y you [gasps] it's like super high pitch >> work pays honest money mostly on time. What have you got? O >> one asset Julian vain CEO of Helix Dynamics. Tomorrow night 8:00 he announces Arya a smart speaker that makes tea and reads bedtime stories. Then you'll appreciate the punch line. Arya doesn't just listen. It harvests banking login records private calls. 20 million homes, including every little old lady within a thousand miles of this shop. [laughter] >> Miz Zealus just pre-ordered one for her grandkids. >> The launch server, the master model, the backups, airgapped in his penthouse, Meridian Tower, no remote, and it needs a physical [laughter] repair man. >> Oh, look. >> Which is why you're not going as a guest. Phone's clean. First class tonight, 11:40 earpiece and the Pegasus drive. Plug it into the master server. It does the rest. 100% up front as always. Then tomorrow, Miss Ellis's bank account belongs to Julian Bank. Wheels up in 3 hours. Say goodbye to the cat. >> Hold down the fort, Pickles. Daddy's got a house call. >> I'm going to I'm going to die before I make it through this. So, this is genuinely better graphically. Aside from the whatever the heck's going on there, this is better than the 5.6 sole result on ultra. Significantly so. All right. Um, don't tell me don't tell me it's bunked out here. I mean, part of me I wouldn't be too upset to not have to sit through the rest of, you know, this extreme. However, I really hope that we Oh, no. Resume. Okay, good. I I'm I pressed E, so we still can't see anything. I'll try one more time. All right. Well, I'm calling it. It gave us enough to be glorious. Um, it just it took this literally took like four hours. I'm not going to give it a chance to fix this right now. It has all the files here. Perhaps this could be something to definitely save. And [clears throat] humor, heart, tension, PG-13, a cozy little felony. Yeah. All right. So, we didn't get past the the initial thing, but it had better graphics overall than 56 Soul. the arm issue. Some of the voices were questionable and um [clears throat] yeah, I don't really have anything to say. That was [laughter] that was hilarious and and it showed promise. All right, so that's going to conclude our first look and test of Kimmy K3. I'm not going to do a traditional results overview for this. The results are in the video. Um it's very good. It really is excellent. This is definitely 100% the best, assuming it does become open weights model. I would need more testing to definitively say if this could go toe-to-toe with GPT5 or 56 with Fable 5, no. But the fact that it's even kind of like still in that as a discussion is extremely impressive and overall it definitely is exciting. It's expensive to use considering the price of previous versions of Kimmy. However, it's significantly larger and the leaping capability definitely does seem to be true.