DeepSeek V4 Flash Is INSANE – The Best Small Model Yet!

summarized

TLDR

DeepSeek V4 Flash is a newly released official version of a small but powerful AI model with 284B parameters (13B active) that achieves impressive benchmark scores and demonstrates strong performance in coding, 3D rendering, and game generation tasks. It is extremely cheap via API and can potentially be run locally on high-end hardware, making it a significant advancement for accessible AI.

Key points

  • DeepSeek V4 Flash has been officially released after a preview period, with significant benchmark improvements.
  • The model shows dramatic improvements on SWE-bench from 7.3 to 54.4, beating GLM 5.2 and approaching Opus 4.8.
  • It is a mixture-of-experts model with 284B total parameters and 13B active, making it feasible for local deployment.
  • API pricing is very low: $0.03 per million input tokens with cache hit, $0.14 without, and $0.28 per million output tokens.
  • The model can be run locally on a 128GB unified memory machine with quantization, e.g., using Unsloth dynamic quants.
  • In tests, it produced a functional browser OS, a skateboarding game, a subway FPS, and a city timeline in C++ without Raylib.
  • It outperforms expectations for its size, rivaling much larger models like Claude Opus 5 in some tasks.
  • The upcoming pro version may be even more competitive, potentially challenging models like Kimmy K3.

Tools mentioned

Techniques

  • Mixture-of-experts architecture
  • Quantization for local deployment
  • C++ game development without Raylib
  • 3D modeling for 3D printing
  • Cinematic game generation with dialogue
Transcript (captions)
It's not funny, but it [clears throat] is. So, Deepseek has released the official version of V4 Flash. Previously, it was in preview. Now, it is no longer in preview. This is a really light announcement post. There's no big blog post or anything. It's basically this X post as well as this benchmark JPEG. And some of the score improvements based off of the preview version are almost making one wonder like, wouldn't they have just called this Deep Seek V4.5? Someone in my discord mentioned that and I agree. So that's why I'm mentioning it from what they mentioned. But this is really like quite quite significant. If we look at this deepsw SWE bench, which a lot of people seem to respect in terms of its actual ability to measure models capability in software tasks, it went from 7.3 to 54.4. It also is being compared against things like GLM 5.2 and Opus 4.8. And apparently it's beating GLM 5.2 and still not at Opus 4.8 level, but this is not a big model. This is flash. This is not the pro version, which is going to be coming soon out of preview as well. And if these gains do persist to the pro version, this may be a very, very interesting early August. So, before we get into it, feel free to subscribe so I can hit that 100K and get the plaque. And someone did ask, if you hit 100K, are you just going to continue asking for like a 500K? No, I'm just going to say instead, like before we get into it, here's my 100K plaque. Like, what's up? Now with that, let's take a look at some of these benchmarks because really there are a few points here I want to drive home about why this is interesting. And the main one is if we swap over to hugging face. Now this has not yet been updated in terms of the open weights version of this which we can see right here. So this is the preview variant we're looking at. However, all things considered it should probably be the same exact size and footprint. So this model is 284 billion parameters. It's a mixture of experts with 13 billion active. That means that this is something that is realistically feasible to be run on a local device. If we go and take a peek at some of the quantizations of this, it is likely going to require like a 2D GX Spark setup, which I understand is still a bit out there in terms of like feasibility of acquiring one of those, but it's more common than you may think in terms of like the local hobbyist world, I guess. So, a 4-bit quantization of this on the low end is like 138 GB in size. If you wanted to go all the way down to one of these unsloth dynamic quants, which we're looking at, you can realistically run this on a single 128 gig unified memory machine. And that's exciting because when this does get updated, it's basically showing if these benchmarks are to be believed in face value that you can have a model that's significantly smaller than these models right here that is actually packing like trading blows with them, which is pretty darn exciting. So, one of the additional things that's linked here in this announcement post is if we click on this, it's just a blog post here that brings us right to how to integrate this with codeex. It seems like this was added by popular demand and is the primary way that they're actually advertising this model. At least when you see that this is the main thing it gets linked to here from the X announcement post, which is basically the announcement post for this model. So for today's test, we're going to be using this just from within codeex. Being that this is a Linux system that we're going to be performing our tests on, there is no pretty codeex desktop guey application, but we're going to be using it with the command line app. It's really not a huge difference aside from maybe just a bit less aesthetically pleasing in terms of watching the interface. But this is all set up properly. They have instructions here on how to ensure that is done. They even have like a simple one-click script to set it up that is recommended. and they show you some of the specific things that get changed here. Now, to note, because it is just going to be saying Deepseek V4 Flash right there, if we go into the models and pricing section, we see that V4 Flash is just going to the newly updated version 0731 for July 31st. We will also be keeping tabs on price because this is really like extremely cheap in terms of model pricing here. So, if we scroll down, it is like is that 2 cents? So, it's like 3 cents per million input tokens with a cash hit. With a cash miss, it's 14 cents per million in and then 28 cents per million out. Though, there are like peak hour multipliers here, I believe. So, you can see that right there. So, during peak hours, it will be twice as expensive. So, 28 cents per million in, 56 cents per million out. So, just keep that in mind as well. So, without wasting any further time, we all know how we're going to begin this, which is of course the browser OS test v2.5, which is the traditional one. Perhaps it's getting close to the time where we need to create the browser OS test v3, which is something I'll probably get to doing sooner than later, just being that this is probably pretty saturated at this point. But it always gives us a nice initial like vibe test of the model if nothing more. All right, it started out so after 7 minutes of thinking, which we didn't get to see anything, it starts by saying this is a fun chunky build. Okay, a weight reference perhaps. no a whole OS, two real 3D games, and it's created this updated plan for itself. So, just based off of what I'm seeing and how long it took to get to this point, I am anticipating this may take quite a bit of time. Now, keep in mind it's on maximum reasoning effort, which is the longest, most verbose reasoning effort, but I always like to test them at full potential if possible. We now have our completed OS, and the first thing I'm a bit curious about is it's named it basically like Ominous. Now, I'm confused about that because we can see that is the name of this system. And also like this is mine. So, it's probably just I'm not trying to sell you anything. Don't worry. These are open sourced. I don't sell them anymore because single person to consumer is terrible. So, that's why I'm a little confused by what it named this. But, nonetheless, we should have it now. So, let's take a peek at our ominous browser OS. This is very, very good, especially for a flash model. I'm definitely now I'm almost like cautious to say these things now because of all like the distillation attacks. This is giving me somewhat of like a GPT5.x aesthetic to it. There is movement to the background. This is probably one of the most busy Notepad apps I've seen. And it just opens with a welcome.txt. Now, let's check the big important thing. Is there a right click? Yeah, there is. Personalize. I think it may have already been personalized. We have a clock in the bottom right that is showing the correct time in the local as well as the correct date. I do like the background movement. It's like two orbs that are just like cycling. We have our start button right there which is just a dot and it has like a floating hover dock. Let's just start by clicking that. Wow. The time machine edition. We can search. So, oh, okay. Do you not have settings? It doesn't. Okay. Let's just run through the apps one by one files. Either this model is going to be like a total freak of nature that's going to scare a lot of people at how potent it is or this prompt has made its way into the training data. I'm going to hold off on any save saved about.txt. Can we actually create something new in this file? Yeah, we can. Yeah. All right. This is very good. Make a new folder, too. It puts it down there. All right. Notepad, which we kind of did already. Just see. It just gives us word count, character count, etc. We need to add a token count to a lot of Notepad apps now because most people are probably just vibe coding. All right. And we can resize this properly. Can we minimize to the taskbar and reopen? Yes. And full screen and unfull. Very good. Next up, GTA City. It's pretty good. Is it quen 3.627 be good? No, but it's close. We have a task as well that we need to start with. I should make sure there's no sound. All right. I have a speaker on, so it doesn't seem like there's sound. Okay. H there's a bit of buggess to this. mainly just like the camera movement is inverted, but we need to steal the gold sports car at the yellow marker, which given the speed we're moving and where the gold sports car seems to be is probably going to take a couple years. So, okay. And it even has like some hover effects around it or something. There is sound. All right. Deliver the car to the green marker. So, it's giving us missions that are basically like stealing cars and then delivering them for money. Solid, right? We even have Yeah, the the [laughter] I'll try to complete this mission and then we'll probably move on. All right, the green marker was down there. Interesting though in terms of like its interactivity. Mission complete. 500 bucks. Nice. Next up, we have our asteroid game, which is pretty common for like Huh. Okay. Shift is to Oh, no. WS is to thrust. Yeah, there definitely seems to be some spatial issues here in terms of Oh, cool. like the controls are not properly mapped to how they should be. Yeah. Nice. And it's split into smaller pieces. That's actually pretty darn cool. Check the big paint test where does the canvas actually scale with this? No, but it it did it correctly enough. Download as PNG terminal. All right. Good looking. Hm. That's quite quite something. Fortune file is a snapshot of your future regrets. That's just like okay. Calculator with the abacus icon. It's like a landscape calculator. At least it scales properly. Good. It maybe should have opened like that. not like that. But I digress. 57 + 6 63 * 5 315 57 + 6 Oh, there was an error. Okay, time machine. This is the special feature. Wow. So, it's auto snapshotting for us. It's concerning how like all the models are suddenly doing this. They're all basically just like recording our actions. It it's timestamped two files and we can restore the snapshot to a specific time. It is really quite interesting how sometimes very different features will suddenly all start to appear across different types of models from different labs and things like that. This is one of those things. Makes you wonder what they're specifically working on behind closed doors. Personalize. Okay. I'm a little concerned. All right, the gradients just aren't showing up in preview, but that one's actually kind of nice. Meadow, it's Windows XP style. Like, it's an ode to the past. Accent color does work. Good. Custom accent color. Oh, all right. Overall, quite good. All right, let's give it the self-contained C++ skateboarding test. This is the retro California aesthetic, and it'll probably take a bit of time. In the meantime, then let's Oh, not that. Let's take a look and see how much that costs. So we were at $9.85. Now we're at $9.77. So the entirety of that cost us about 8. That is exceptionally well done for a model of this size. Okay, so we can't ollie. Seems like the tricks are not working 100%, but this is much better than a model of this size would produce by a very, very high margin. The pedestrians look good. Even if the tricks are not actually specifically working right now, this is very, very well done. I'm going to give it some feedback based off of this. I think that was a start screen that we just saw where the text was definitely visible now, but it was perhaps a bit scaled up. No image viewer, so we'll verify the heads-up display by analyzing pixel content and rendering the region as AS asy. All right, let's take a look at our improved skate game. Okay, so it did get the text kind of working. It's all right. [laughter] Better than it was good. Now the ollie works. It fixed the main issues. Look at that. [laughter] What is Okay, that's is that's not quite okay. Hold on. So, the kick flip is K, which I guess now saying that out loud makes sense. L is some form of oh, 360 flip. So, I'm going to say it actually did an okay job with the text. The thing is like most times a model of this size would just like mangle the text pretty badly. This actually made a significant amount of progress. I mean, okay, maybe now that we see that it's somewhat impeding our ability to see, but the shops do now have their signs on them. Can we snap to the rail and grind? There may be some like specific key that needs to be pressed to snap to the rail. Sometimes that is the case. But I mean, regardless, yeah, look at that balance, too. What other model did that? I think it was Gemini 3.6 Flash that did balance when on the rail, and that's not something I've really seen before. So, this is overall significantly better than you would expect to get from a flash model. I mean, again, if this isn't benchmarks, this pro model might absolutely like just come out and decimate every other current model, which would be exciting for the consumer. All right, we're on the beach now. Can we go in the water? Surf sim. Okay. Yeah, we can. That's [laughter] It looks like a surf simulator now with the arms like flapping. It's Yeah, I'm not gonna say that. All right. Overall, this was sick. The 360 flip is sick. Look at that. Wow. I mean, that's pretty darn good. So, next up, we're going to be giving it the beautiful static subway scene test. And this is the prompt that includes both just making a subway scene in 3D that looks good, as well as turning it into an FPS. So, we'll see what we get. All right, let's check our Subway FPS result. Subterror. Okay, explore station or star combat. It's interesting when I give them the unified prompt. Often times we get the option to either just explore or play an FPS. That's a Ban issue. Yeah, that that'll fix that. All right, we have some sound here. It's incredibly [music] [laughter] okay. It's incredibly trigger happy, I should say. Again, for the size of this model, this is just like really good. We have some footsteps. I don't see any enemies, which is making it creepier. We're on wave zero, so I I I guess that B is for bloom. Wow, that's just nice to have. F is flashlight. Look at that. So, really the big issue here is there are no enemies. Okay, [music] that's that's a that's a problem, but you know, we'll work around that. Wave zero. So, it seems like we have some issues here just in terms of gameplay logic. Let's just do explore station. Cool. And we basically have the same exact thing. Sands the weapon, which makes a bit more sense. All right. Hypothetically, we have our fixed like combat issue. So, the autofiring weapon should be no more. Um, [music] okay. Well, that is We'll just ignore that. We do have enemies now. Let's take a look at them. Okay. Oh, hey, it's not very nice. [laughter] >> [laughter] >> That is I'm just going to be straight up like [laughter] what? All right. That's just watch. It's not funny, but it is. [laughter] This model's sick, too. It's like Opus five levels of sick. All right. Well, I've seen enough of this. That was quite something. [laughter] All right. I want to try this with something that I've just never done before, ever with any model. It's going to be a C++ game, but it needs to be an old school looking 3D motorcycle game. Because this model's been performing so well in some of these prompts that have been run before with various models, I want to see if that performance translates into something that I've genuinely never run before, ever with any model. So, this is going to culminate in hopefully a pretty decent C++ self-contained retro motorcycle racing game. And if this looks very impressive as well, then this is really quite something. All right, so it's been working on our Moto GP game, which looks pretty good. The thing is it used ray liib. I didn't catch it in time to tell it no, you need to do it like in hard mode. So that's unfortunately going to kind of skew the results here because the whole reason I'm running this is to see how this does on a prompt that has never been used before. So it's a better generalization of the raw capability. So I don't quite know what to do. I mean we'll take a quick peek at it. I mean that looks pretty darn good. Moto GP the that's very good. The thing is it made it a bit simpler by using ray. So I don't know 100%. The handles I think are inverted. It seems to have some issues with like spatial. I'm just saying don't use ray. I spelled it wrong. Start over. We'll just see. It's possible that it may still have like things cached from the history of what it just did. But nonetheless, we'll just see what we get. All right. Here's our Moto GP game test where it can't use Ray Lib. No, this I'm afraid this test has been ruined because I didn't catch it when it was using Ray first. So, this result is entirely Sans Ray and it's actually better than the one that used Ray, which is somewhat perplexing. It's very, very, very good. I mean, interesting. All right, I'm giving this the beautiful website for Slapis Watch Company. It needs to make a good-looking front end for a website that has a 3D model of a watch, cinematic panning camera view on it, and then just like a decent competent website. All right, let's take a look at our beautiful watch website. So, it implemented like a render feature time distilled. Nice. That's a good good bit of lingo to use for a website that's this suspiciously clean now. Um, okay. So, we're going to notice there are some blatant issues here. I'm noticing a common theme in where the results are weak with this model is that spatially it just doesn't have a great grasp on like proper orientation in 3D space, I guess, as how we would view things. But, if we were to judge some of these elements individually, the wood is very nice. Perhaps a bit too small for this watch. The actual materials look pretty good. We can somewhat move around, but unfortunately I can't like really take control of this orbit. It has the elements there. They're just not necessarily arranged in the proper orientation. But still, I mean, okay, that's actually So, it rendered this online and then took photos of it. When I say online, it spun up like a Chrome instance to do some renders and then take photos of them and then include them here. That's what we're seeing right here. Is this supposed to be a pocket watch? I don't think so. Okay, it just kind of was mangled. But look, it even has like sub dials and slapis. This one looks pretty good. Look, man. This one's actually better. Well, it's messed up, but like Lagoon Diver 300. All right. How much are these bad boys? Not bad. Very, very affordable for your average influencer. No, I'm kidding. [laughter] I'm kidding. But this is uh pretty interesting. All right. Overall, this was still very cool. even if it had some spatial issues. Next up, I'd like to try a 3D model test, but instead of the traditional V8 one, I'm going to change things up slightly. So, I need a model of an inline 6 engine. It must fit a 280 size DC motor inside of it for my RC car project. It needs to be able to be 3D printed with minimal to no support and properly house the DC motor. An inline 6 engine is basically just like a a rectangle like that, whereas a V8 is a V with like different stuff. So this is just a bit of differentiientation to try to test the capabilities outside of the more traditional prompts. All right, we have our 3D model of the inline 6 done. There are some renders there. Let's just open it in OpenCAD. Okay, it's So here's the thing. This is not our traditional V8 test. This is an inline 6, which is correct for what this should look like. like from top down definitely. It has like an oil cap there. The big problem I'm noticing here is the render of this 280 DC motor is just completely oriented in the wrong way. But this does seem to have some like world space oddities to it. However, this could work depending on like if there's a hole in this cap for the motor. Let's look at some of the actual like print files. So, here's okay. Body.stl. All right. It's big. [laughter] Okay. So, does that have a hole for the shaft? It does have a shaft hole, which is good. Additionally, it even has like outlets here for um like heat dissipation wires, whatever you want to refer to that. Son of a gun. And it did the like V-shaped thing. So, it won't need supports at all. We probably could have had a circle down there and then just a V at the top, but I think Kimmy K3 did this as well where it had a V angle at the top when we did this with the V8. It does. Okay. So, if I were to nitpick, yeah, like these walls are not thick enough to properly accommodate these holes. So, they should either be closer to there or they should just have like thicker wall out there, but there should also be then the rear cover would like slide over the motor and probably just bolt in. Very good. Okay. Um, this ring is maybe not 100%, but still just the fact that it did this in multiple pieces. Then we have the crank pulley. All right. Very cool. and just for more detail if nothing more for the motor. And then we had the fit fit check. Isn't that like a thing? Like like a fit check would be like how's my shirt look or something? I don't know. But and then Okay, cool. So, it just made a little model of the motor to properly be able to fit it in to the body. So, still this is I think it's actually quite all right and quite well done. It seems like it would actually be realistically printable and the motor would fit in and it wouldn't be like a big huge mess of like different non-connected bodies. So very very solid. Get it? Cuz it's a solid button up. All right. This thing seems really good. So I'm just going to go hard mode here. This is kind of almost messed up to ask this to do, but I'm giving it the Chrono City test, which shows the city in different time periods, but I'm telling it to do it in C++ without using Rayb. So we'll see what we get. This is going to be like this is messed up, but it's a good test that's out of scope. This C++ city timeline result is taking quite a while. So the point we're at is like the UI and things are loading, but it's just consistently a black screen. We don't actually see anything beyond just the control map as well as the timeline slider. So we'll see what we get. All right, that's that was pretty good. That was the first I've seen of this and I noticed a few things that were problematic such as like sideways mangled cars, but this does seem to have like somewhat of a look at the transition between the time. Keep in mind this is all done using C++ and we're not allowing it to use Ray Lib or anything. So this is a difficult test, something I've never done before. We've done Chrono City, but it usually, not usually, it always has to be an HTML file. This is C++. So I mean, [laughter] Yep. All right. So that's 2055. It spoiled like a lot of the wow for us, but the fact that this has actually pulled this off is pretty darn awesome, especially considering the model size. I know I'm going to keep driving that home, but still that's a big thing because say you have two of these DGX Sparks, which I probably will very soon. But I mean, for $9,000, you can run this locally with probably a million context with Yes, that's a lot of money. But like 6 years ago, if you knew you could run something with this performance locally for $9,000, it would be like this is magic. So that's why I'm so excited about this. It's also very cheap to use in the API. All right. So, it even has like music of the time. Can we move around? Yep. [sighs] Oh, it's I see little people. [laughter] It's automatically swapping through the timeline for us. All right. There's some issues here, but like it's still This isn't C++. QE is down and up. C is a cinematic tour. Ah, okay. Maybe maybe we need to undo this cinematic tour, please. Okay, there we go. All right. How do I pause the space? It's not made it easy for us to really look at this, but like this is still incredibly good even with the issues. We see people here. Okay, so this is 1945. Let's see if we can go down further. There we go. That's that's a little better. If it had kind of brought us in here. Okay, the movement's a bit odd and the cars are obviously mangled, but this has some problems with like its spatial understanding. Let's go now to Okay, two. We have 1965. The car is a nicer color. Some of the signs have changed. More teal, like like pastel colors. The people more or less look the same. Three would be 1985. Okay. The music changes. It's, you know, I think we had street lights that changed, did they? Uh, kind of. That's okay. Four would be tail fins, neon ribbons, concrete, synth, base, glass towers, and screens. Oh, why did it go to 2005? All right, there's 85. Now, 2005. The way the buildings are moving is very cool. It looks like something out of Harry Potter when they go to the secret Order of the Phoenix hideout. And then finally, the future, which is filled with blue trees for some reason. I I mean, I'm down with that. I like that. Blue is my favorite color. And very, very interesting. So, still, this was impressive and something that we've never at all done before because this was a very, very difficult task. And the music is well done as [music] well. So, I mean, yeah. All right, let's check our total cost for all of the tests we've run so far. $96. So we started at $9.85. So 79 for the entire duration of everything we've done. Now that is including something additional that I have not yet talked about. So for the entirety basically of this test from within open code. I have given this the Steve the PC repairman cinematic game test. This is something that I give to things like Fable 5 or Kimmy K3, not smaller models like this that are not gigantic state-of-the-art Frontier models. This is a difficult test because it basically needs to make the full gameplay. I've given it the open AI API key to generate dialogue. It has to do cutscenes, cinematic camera pan, story line, etc. Now, this has been going on for quite quite a long time, and I just want to see what we have, if nothing more, just to get some form of like, you know. All right, the sound isn't working. Damn, look at that. That's not bad for the game. So, again, it's not like I'm not expecting perfection here. The tab is muted, isn't it? >> That's when we bought >> I suck when it comes to like speaker handling. And >> Tuesday, 5:47 p.m. Steve's PC repair, where the elderly are never on. >> Oh, you can't see what I'm pointing at. I'm sorry. Are >> you sure she'll live, Stephen? She'll be fine. Bizelis, it's just the CMOS battery. >> 1998 was a very good year. That's when we bought >> when we bought the Buick. That is something an old person would do in that year. And I'm not being like, you know, I have a Buick myself, so I'm not being like mean. All right. Replace the CMOS battery in the computer. How would I go about doing that? Enter. E. They usually have like similar like key mappings. Nice. Unscrew the side panel. E. Pop out the old CMOS battery. Drop in the fresh battery. Screw the panel back on. >> There we go. New battery. Case closed. >> Is that the >> Oh, it's glowing. >> These people are >> carry her out to the car. >> Troubled. >> Oh, I have to walk this lady out. I mean, that's fine. You know, it's what a good store owner would do. Tommy watches you work always. Okay. Uh, where's the door? So we can vending machine. The machine demands coin. Old school. You respect it. Rules of the shop. Okay. How do I get out of this shop? Again, some spatial issues here. Okay. Miss Alice is leaving. These sounds are like not entirely EQed properly. So, oh, there's a door. Okay, good. Here's Oleg. This is all doing it itself. He nods. It means a lot. Okay, see what Mr. Thomas wants. >> Mr. Thomas busy. Ms. Ellis's be. >> It's kind of like is auctioning off nemesis. [laughter] Black most not stolen you. Everything except a first class ticket. It's doing it too fast. Like >> I'm in. >> Open the briefcase. The fact that this is like actually Is that the briefcase? Yeah. >> Phone lockpick. >> PDU KT. Nice >> key card gets you to floor 40. The hard part is above that. >> Take the [music] phone lockpicks USB worm key card. Do I have to Okay, now what do I do? Fly to Helix Tower. Go out the front door. Two hours [music] later, above the clouds, the cats sitter texted. Tommy says, "Hi." The peanuts are actually good up here. I think we're supposed to see like in the plane view from the airplane. I mean, >> if it taps out here, that's fine cuz like Helix Tower 9:47 p.m. >> What the Look at those like floating. Do I have to I have to hide from security, don't I? >> [sighs] >> If this Nope. I'm sorry. Oh, I have to, don't I? >> Okay. I can't. I'm sorry. It was to be honest with spaces to skip. Yeah, I was doing a lot of that and I can't get it to be gets quite frustrated with games. I don't play much video game at all, so I'm not one that like enjoys the struggle. All right, I don't sorry, but like no, I'm vetoing it. That was still significantly better than what I would have expected by a long long shot. I mean meta's muse spark 1.1 I gave it this test and it was like this did better than that and that is not what I expected. So this is a more difficult test that like Fable 5 Kimmy K3 did really well on. GPT56 Soul didn't even do that good. So this is like just showing some real proper capability here and that's my culmination of the entirety of our testing today. This model seems truly truly very impressive to the point where we ran it on a few things that were not ever done on the channel because I was a bit almost suspicious of the performance it was putting out for the size. I don't normally get so like suspicious when testing these models. I just see how they do in the tests and it's like, oh, okay. But this one was just so good for the size. And a big part of the reason that I'm really excited about this is again this is something that actually has a realistic possibility of running locally on people's systems. Yes, you're going to need a pretty expensive hobbyist machine, but considering the performance you could be able to get from this, the cost is not entirely unjustified for those who can afford it, who want like local AI, and that is still a lot of folks even if it's a niche market for now. again the the 3D model the subway FPS which was just disgustingly sick and something like Claude Opus 5 would put out because that model is also like a sick in the head model. This one also seems to have some of that like sickness as well shared with it. And the way we could turn the bloom on and off for the station like this was just really like Yep. And it did have bullet holes in the environment. I didn't even notice that to comment on it because I was so busy just being like, "Yo, this model's really good." We have the flashlight capability as well. I mean, that's awesome. And then the biggest issue here was the gun would just auto fire for us. Then it would leave this like just disgustingness in the station. But still overall just like properly properly competent for a model of this size. I want to keep driving that home because this really is quite awesome. I mean, the skate game was really good. And again, this was one that made me like, okay, I want to do a different C++ test as well, which is why we did both the Moto GP one, which was good, but it did it first in Ray Lib before I could catch it. Then I deleted everything it made and said do it without ray, and it still did a good job, but it basically looked like exactly the same. So, I don't know what sort of memory of that it had. And then also there was of course the city timeline one which was just like quite competent for being done in C++ without really it made the music and things like that. So overall I think a lot of folks are going to be extremely extremely excited about this model because it's very cheap. It's locally hostable. And if these gains actually translate to the big pro model, that could really be a Kimmy K3 competitor, which I'm excited about because it's significantly smaller. That's a 1.6T versus a 2.8T. So, pretty interesting. And that's probably going to conclude today's video. So, if you have any questions, leave them in the comments. And thanks for watching.

Frontier News · by Hyperjump Technology