DeepSeek V4.1 Flash Is INSANE – Is THIS the Best Open Model Yet?

0:00 Okay. Yeah, I'm the the big red one. Heat not ready. Deepseek has introduced V4.1 flash, which is honestly something that I think could have been called V4.5. This is not

0:16 just an iterative improvement over the previous flash model. This is basically an entire new model. It has a new architecture, a new size, a pretty cheap price, and some based on these benchmark

0:26 JPEGs, pretty potent performance. So before we get into it, do feel free to subscribe so we can get that 75K and then hopefully the 100K. And let's start out by taking a look at the newly

0:36 releasleased V4.1 Flash, which really does seem like it may be quite a significant interesting release, I guess could be said. So starting out right here, let's just look overall at these

0:46 benchmark JPEGs. Now, I don't really put much weight on these at all anymore, just based off of some recent scores. However, it does seem that this is comparable to something like Kimmy K3,

0:56 which is a 2.8 8 trillion parameter model significantly larger than V4.1 flash as well as on par with GLM 53 Opus 5 and 56 soul. Now something interesting about this model and while we are mainly

1:09 going to be doing our traditional testing here where we're not focused as much on some of the technicals but more on the output, it is something to make note of right here that this is a 552

1:19 billion parameter mixture of experts model. It has an encoder decoder architecture with 8 billion active parameters per input and 16 billion per output. Now, don't quote me on this, but

1:30 there's something here called like a T5, which is an encoder decoder transformer available in these size ranges, etc. Something fairly old. I believe in 2019 Google came out with this. It's just

1:41 something to make note of. If you hear this being referenced as well with this new model, there's just maybe some additional reading here and you could paste it into an AI and say like, how

1:49 are these things similar? um which would do a better job than I would of trying to make sense of that. Now, with that as well, we have some more verbose benchmarks right here. And again, these

1:58 I'm not as keen on since due to recent events with Deep S. However, I find that this does seem like it stacks up pretty well and really the main and important thing is it seems this is really

2:09 outperforming V4 Pro even though this is a smaller model. So, that's pretty exciting to see. This is the smallest model in their new architecture family. So that basically means we can expect to

2:19 see something with a lot more parameters of this architecture which would then hypothetically scale these scores much higher. Additionally, something else that is really going to be I think

2:29 pretty impressive when folks wake up in like other parts of the world. The massive reduction in KV cache that's being shown right here. This is to put it simply basically means that you can

2:40 handle much longer context without needing as much available RAM basically. And that's even what they're showing right here is how much resource is used right here for this level of KV cache.

2:51 And we can see the reductions right here have been getting significantly smaller. So this bodess well just for being able to handle long context without needing a ton of extra resource I guess could be

3:02 said. Now I know I'm going through this a bit quick but I am really excited just to test it and also um yeah probably just that we can see right here that the model is available on the API right now.

3:13 Something else to make note of is that this is natively multimodal. It can see images, something that the previous DeepSeek models up until the vision experimental ones. So, pricing for this

3:23 model does seem to be pretty darn good, assuming the benchmarks hold true. At the most expensive potential use case, this is 30 cents per million input tokens and $1.20 per million out. And

3:34 the model weights have been put on HuggingFace, so it's not one of those like, okay, we're going to release this later. It's just here now under the MIT license. Now, something to make note of

3:43 is the previous Deepseek Flash model. I don't actually remember off the top of my head how large it was. Okay, so it was something that could be hypothetically run on like two DGX

3:54 Sparks. This perhaps not. However, there is some hope there as well. So, if we actually look specifically, and I have it open here as well. If we actually go in and look at these specific files, if

4:04 we go to the end, we're going to see that these two are gigantic compared to all of the other parts of this model right here. Reason being, these are basically engrams. So to put it simply,

4:14 and again, don't quote me on this unscientifically, these are things that can actually realistically be offloaded to CPU memory or perhaps even streamed from a disc if you had a speedy SSD. So

4:25 while this is still a pretty large model, it does have and somewhere here they mention it. There it is. Good. So I took me a while to find it. So 196 billion of these

4:37 parameters are these engram parameters. So again, these are things that could hypothetically be stored or not hypothetically. You can basically offload those at least to CPU RAM. So

4:48 you don't have to worry about the entirety of this being shoved on a GPU. I don't know how this will work uh running locally is what I want to say. But the entire size here is not all just

4:57 like I need a beefy GPU to run this. So keep that in mind as well. There is some hope maybe of some magic being done to get this feasible on still expensive but hobbyist systems. So, let's quickly take

5:08 a look at how we're going to be testing this today. First off, I do want to make note that we will be monitoring the cost for each of these tasks right here. Currently, I have $312 in here. Now, if

5:18 I refresh this right now, that should go down to something. So, I have already run the browser OS test v2.7. That cost 25. And the way we're going to be testing this is using the deepse

5:29 harness. I find it works pretty well. And something I would like to make note of is I mistakenly ran this on high effort mode instead of max. I suppose for the browser OS it's probably okay to

5:40 do that being that every other test we run will be on max and probably perhaps a bit more difficult warranting that extra level of thought. So with that we do have our browser OS right here. It

5:51 cost 25 cents on the dot and let's take a look at it. So let's take a look at our browser OS run on high. Okay, this actually looks pretty darn good as well as very very very similar to something

6:03 I've seen before. Just probably I suppose that makes sense. Let's see. This almost looks like a dead ringer for I think the previous I don't remember, but something I've made looks exactly

6:15 like this. I guess should be said. Okay, Harbor Edition 11 resident programs. Two of them are 3D. One of them rewinds time. We don't have a right click, but we do have the clock and the date in the

6:25 bottom right, as well as our audio settings, our simulated network settings, and then our chrono rewind snapshots. Wow. And this is just taking snapshots every That's a two second gap.

6:37 That's a two-cond gap. That's a 7-second gap. Maybe when Windows states have been changed, it would probably make more sense. Okay, so now with that, let's start just with our

6:50 start menu. Harbor Mail. Why not? I know we're going a little out of the way. Another name, Voss, in these. That's a pretty common name for AI. If you used AI for any sort

7:01 of writing at all, you very likely seen something Voss in one of the names. Okay. And as with this test, generally all of the emails are related to apps on this specific system. Now, with that,

7:12 let's turn our speaker on and try our GTA clone. And remember, this was on high, not max. So there still would be something to go. Okay, this is the potential here is

7:26 is high. Interesting. I've not seen this many trees in a GTA game. Now, yes, we're going to notice some things are not quite right. However, that could be on me for neglecting to run it on max.

7:38 Let's just see how the second game is. Oh, this is I don't know. Like, oh, is this us? Okay. I think this white pyramid ship

7:55 thing is us. Interesting. And one of the options when we lose is just new tunnel seed. So, there's like a randomization factor, I guess. Interesting.

8:08 Okay, we looked at our mail thing and the browser OS. I still like to do them just because it's like nostalgic at this point. However, they're not going to really show us a huge level of

8:18 capability at this point with these models and how they are becoming so much more powerful. Interesting. Look at that. The music file actually has a specific BPM, scale, and voices. I

8:29 wonder if there's some form of music app in here. Terminal crew. Okay. Weird username. Jukebox. Good. It does have procedural audio. And we can

8:56 uhoh new melody. Okay, so that does actually work. A nice little surprise as I do like when there are audio related things. I feel like we had one or two more apps here. Perhaps

9:10 we didn't look at wallpaper. I like that. Motion speed settings. Foreground particles. Constellation lattice.

9:25 Oo, plasma field. Oh, I like this one. Topographic drift. Yeah, that's actually pretty cool. That's a pretty unique one. I've not really seen this style before. Hyperspace. These are actually really

9:37 nicely done. Oh, wow. This has a lot of wallpaper customizability built in which is pretty cool. Open timeline. And then we have our snapshots just of different windows being open. 50 moments held.

9:51 That's pretty cool. Overall, this was pretty nicely done, especially considering the fact that it was not run on the highest effort mode. Now, in lie of trying to get some of these things

10:00 fixed in terms of like the GTA game and things like that, I want to just run more intricate tests that will showcase capability more so than just the built-in GTA clone. But we can see right

10:10 here, it does have walking animations. There were vehicles and things of the sort. So, even if there were some rendering glitches, overall it did actually have some level of

10:17 functionality and some movement. So, I'm happy with that so far. All right, we're now on max effort mode and we're going to give this the self-contained C++ skateboard game. This is the one where

10:27 it needs to make it a New York City block as the environment it's skating around in. And we can see right here the speed is just absolutely in line with being named Flash. All right, this has

10:37 been going for quite a while and it does seem like it's about to wrap up. Now, I had to actually stop this for a moment because I noticed it was trying to look at images of the game and it was like

10:47 this model doesn't accept images. So, I realized there was a leftover configuration issue in this harness. So, I just made it understand, no, you can't see images and continued it. So there

10:57 that wouldn't have really paused anything, but just in case you see like the two turns right there, that's why. So we now have the specific game and the command to run it. Let's just take a

11:06 look and see how this did. Okay, enter start two minute run. F is for freekate. 1 to four warp to a spot. W to push and to steer. Okay. Beep beep. All right. This is

11:24 this is really good. And I say that for a couple of reasons. Well, okay, it looks like these NPCs are like they're like power walking. I think the actual movement of the rider

11:38 is pretty good. These NPC this has some really really impressive NPC density. Can we interact with them? We can. Okay, so that I like to see. The amount of detail here is actually pretty good. It

11:50 doesn't seem like we can interact with the cars, but like we have things like the puddle, we have the steam there. The cars actually honk at us just we can see the beep beep. The tricks here though

12:00 and the movement there's a pigeon are pretty good. The fire hydrant shooting water out. Something that like every model that can competently do this includes. It's just interesting to see

12:10 what the models assume a New York City block looks like apparently. Okay, we have this pedestrian like moonwalking across the crosswalk and is about to get run over.

12:22 These pedestrians are funky. But this is a this is a competent result, especially when you consider the cost that it took for this to basically make this. So, which we don't actually have

12:36 yet. We'll take a look at that. Oh, that's cool. Okay, cuz it had those hot keys where we can kind of jump into different locations. I really want to get in that halfpipe, but I don't know

12:47 if we can cuz it seems closed off from both sides. Yes. All right, let's see what's up here. Let's check like the actual like physics, I guess. Okay, it's not leaning like it should be.

13:01 However, when we go up and we let go and we roll back down, that's decent. Let's try to do some like X games. Oh, okay. All right. And that's like a zombieoid looking person. The bail effects are

13:14 pretty harsh. Get out of the All right. And it only gives you two minutes. That's actually pretty interesting. Oh, F is for freekate. Okay.

13:26 Okay. Kind of. Yeah, you can actually see the board tilt. It's just hard to And the board is actually pretty bad. I didn't notice that before cuz I was enamored with the rest of it. The board

13:39 is, you know, not the best, but overall, this is a nicely done and playable result. Can we grind? Yes. Okay, that was a cool angle. All right, I

13:54 enjoyed this. This, I think, is pretty well done, especially for a Flash model, although it's a pretty beefy model. This next test is something that involves controlling a physical robot based off

14:04 of what it can see from an attached webcam where it will essentially have some information about the robotic arm it's controlling. It will have a camera placed in a suboptimal location and then

14:14 a task basically use this robotic arm to pick up the orange car and then move it to the other side of the platform. So unfortunately after about 45 minutes I opted to stop this. One of the big

14:25 issues I was noticing was that it did not seem to understand the grippers were inverted. So when it would reach down, it assumed they were open and then it would slowly open them, assuming that

14:36 was closing them. It had a lot of trouble. And I would just assess overall I think this was a vision capability issue and not necessarily a model intelligence issue. So I think that with

14:47 a better vision pipeline, this may have been able to get it. It was still interesting to see, but it just wasn't quite getting it. And the inverted jaws kind of pushed me over the edge to be

14:59 like, "Okay, I think 45 minutes is enough time here." So, unfortunately, that didn't go too well. And I did opt to stop it basically after I think about 45 minutes. Now, something interesting,

15:09 and I don't want this to be considered like a total loss of a test, is that initially when I had started this, it actually was having an intermittent connection issue with the USB plugged

15:19 into the arm. So, it interestingly noticed that the cable was bad. I swapped it, but it was like, okay, well, it has a Bluetooth receiver in it. Like, let me try to get it that. And that was

15:28 very interesting. And then the camera it was using, it was like, I'd like a fixed exposure. And then it was like, oh, this camera doesn't support that. So, let me understand what it does. It seemed to

15:38 show some pretty potent like software reverse engineering capabilities here just in this test and trying to figure out exactly how to get the arm working and things like this. So that was

15:49 interesting to see. It's plugged in now. Bad cable I had sent to it at 458. So, and then it figured out, oh, okay. And it moved on from there. So, really, unfortunately, the big issue, I presume,

16:00 would be the vision was letting it down. We do have a folder here of a large large amount of captures that it had created. So, we can see right here what it was doing with bounding boxes,

16:10 zooming in, and things like this. Now, there are many, many of these on here, and we can see all the work that it did. So, even if it didn't get a visually exciting result in terms of actually

16:20 being able to get the car, it still did do a bit of work. So, it's interesting right here to even see. And you can see from these images, it could be confusing. Like, wait, the grippers on

16:30 the car, like especially this one right here, like how is it not grabbing it right here? And that is why I do this test in this specific way because it's very difficult. If there was another

16:40 camera angle, it would be pretty easy to do like a AB C like by depth perception thing, but it's difficult for a reason here. And I like testing the models with more difficult things just like this.

16:54 This was when it was trying to figure out exposures and things like that. I mean, look at all the work it did. So, it's interesting nonetheless. Even if the result is not successful in terms of

17:03 the car getting picked up and moved, it still shows us some of the capability. And I think that was pretty cool just to see regardless. So interesting. Next up, we're going to be giving it the Subway

17:13 FPS test. I need something like fun and light-hearted after that test cuz it was kind of stressful sitting there for 45 minutes watching it being like the grippers are inverted. So, um, this will

17:23 be a nice, you know, palette cleanser, so to speak. All right, I'm going to be totally honest. I've been kind of disappointed with the Subway FPS results. So, first and foremost, we see

17:32 this is heavily inspired seemingly by the London Underground symbol, and it's called Ridley Road. Now, the biggest issue I had was when this first produced its result, and the Deep Sea Harness was

17:43 kind of freezing up a bit, so I had to like spam click stop and stuff cuz I was worried about the screen recording that was ongoing. Basically, first off, we didn't have any weapon models and also

17:54 there was no ability to shoot perhaps due to the fact that there were no weapon models. Now, this is the updated prompt where we also have a train that provides us with the subsequent waves of

18:04 enemies. Okay, good. Okay, now I'm starting to I'm starting to like what I see. The actual ammunition seems to be reflective of the specific weapon. It's very um kind of like dork and these

18:18 floating stains in midair. The next train is not scheduled. We apologize for the delay. Look at the slide of this pistol. Okay, so it's totally just frozen up now. And that's kind of one of

18:29 the issues here is Oh, okay. Okay. Well, ignore the locomotive car and maybe the actual train car is not bad.

18:42 The doors open. Good. And we do see them actually coming out. I will say that. Oh, are there holes in the environment? Yeah, there are. This has potential, but truthfully,

18:56 it's a bit more troubled than I would have liked it to have been. It's possible that my judgment on that is based off of the fact that the harness was kind of glitching out and the

19:06 computer doesn't seem to be too happy right now. So, I mean, this right here is well, it looks like I'm blurring something out on the screen and I'm not. So, it's got potential, but all right,

19:17 we've got a bit of sound now. So, let's just take a look at this quickly. Um, I do want to just try the There must still be an enemy. Yep.

19:34 Yeah, this isn't bad. I just I think it could have done better. Hey, the reload's actually kind of more in line with realism. Can we shoot the lights out? No.

19:47 Some of the detail here actually on the track and things like that is is not half bad. Apparently there's one enemy left and I can't actually locate them. So, but we did see the train effects

19:58 before. Now, I'm going to check usage right here. Keep in mind, okay, we're running pretty low, but that is definitely justified because additionally to our Subway game for over

20:07 the past hour, it has been working on a GDAU result using Blender as well for assets, which is, if I go back to Chrome, the 1980s 3D wrestling game prompt told to use GDAU and Blender to

20:18 create this. It's been working quite a while here, and I am pretty darn excited to see what we get because it'll showcase some of its 3D modeling capabilities as well as using something

20:27 aside from 3JS. does seem like it's getting closer to finishing just based off of some of the things I was seeing right here. Excellent. This is a huge improvement. I'm so tempted to look at

20:37 one of the images. I'm not going to, but it's extremely tempting. So, we'll just wait for it to finish. And this goal is ongoing as we can see right here. All right. We finally have our wrestling

20:46 game finished. Now, I did top this back up in terms of the balance because it had dropped down to like less than 30. Basically, just so far for the entirety of the testing we've done, it's been

20:56 about $2.80. Keep in mind there is currently ongoing a slapis watch website result. So that's attributed to that value as well. With that, let's just check this GDAU game. I am quite eager

21:08 to actually see this. I think it may be pretty fun. Hopefully. All right. So far, this looks very good. And I noticed when it was working, it was doing a lot of like rigging the animations and the

21:21 characters and things like that. So, okay. Uh maybe we'll just try to run that once more. Oh, okay. The audio it like made a lot. It really worked very very hard on

21:40 this. I will say I'm going to at least for the duration of this turn the music off. Let's just select a wrestler. Oh wow. Okay. Interesting. Um, he will break you.

21:57 Moscow USSR 68319. Okay, we can select. We can Okay, the Neon Express Johnny Vega. Look at that. That's actually a decent model. The problem is the arms are coming in

22:11 and out of the actual outfits that they're placed in. Something else I'm seeing that concerns me that happens a lot of times with these games is the wrestlers fall through the floor. Um,

22:30 I can't show this because the wrestlers need to have shirts on from the Blender model assets for me to be able to show this properly. So, I mean, they did in the preview, but then

22:44 when we jumped in the game, that was not the case. So on it, I'll add a fitted colored shirt to each wrestler and rebuild the GLBs. All right. So, there have been some

23:05 issues with the wrestler models that this created in Blender. So, the big thing is I can't show them. Essentially, I'm having it do a quick follow-up where

23:16 they all need to have shirts on. Now, with that though, we can at least get get a preview of some of what it's done here with one of the wrestlers, Vega. Um, can I Oh, well. Okay, so here's Vega

23:29 in his outfit. Let's just I'm just going to hit start match. Oh, thank goodness. Okay, so this is much better. Now, this actually did a interesting job. Unfortunately, they're

23:49 not as rigged as it seemed like it intended them to be. It was talking about many different movable joints and things like that. Also, the VCR glitch effect or something is is kind of

23:59 detracting from the game, but I'll say in terms of the actual Soviet slam is ready. Oh, okay. I Okay. Yeah, I'm the the big red one. Heat not ready. Okay. Well, I'm trying

24:17 to hit the other people, but Oh, cool. Okay. So, like that right there looks cool. Aside from the the glitching is just upsetting. Let's just pick someone and All right.

24:46 I'm done. I'm done with this. I'm done with this result. All right. Regrettably, I I feel like this is worthy of a additional look just because some of the things here were actually

24:56 quite impressive. Things like the random bits of audience that would get highlighted as if there was like flash or a strobe light. The actual environment itself outside of the ring.

25:06 This is actually, I think, a pretty impressive result when considering that it made this with Blender and GDAU. Even the ring itself looks pretty good. So, I'm gonna just do another start match

25:18 here. And maybe we'll just keep in mind like, and it it even like rigged them in some different ways. You can actually see slight movements of individual parts of

25:29 their hands and arms and things like that. So, it's it's definitely got potential. The weird glitching effect is unfortunate as that's kind of caused some

25:48 issues. It's all right, let's just All right, I'm having trouble even like moving right here.

26:05 Lone store DTT is ready. P. Oh, I I pressed it, but it doesn't seem I'm having a tough time actually getting the finishers to work. But overall, if we actually look at this, aside from the

26:16 hilarity of some of the things here, it seems like it's actually kind of impress kind of impressive. All right. Oh, okay. Good.

26:29 and we're moving around. So, I would assume they're going to come back in the ring. Yeah, it's just it's a bit troubled, but I definitely think this was worthy of a

26:41 a look once I kind of gathered myself after the initial hilarity. Okay, so you know, definitely interesting. So,

26:56 while I was gathering myself after that wrestling result, I gave this another C++ test because I like to do multiples of these if time allows. In this case, this is a retro 3D rally game that it

27:07 needed to create with C++. A firsterson view from inside the car, kind of a small interior of the vehicle, low poly, reminiscent of early rally games. And I've also instructed it not to use Rayb.

27:18 It said right here when it gave us the culmination of everything it did. I believe it said over 7,000 lines of C++ across 30 files in this script. So I just wiped the folder that the um robot

27:30 arm job was done in. So we should have our result right here. And I'm actually pretty interested to see this. That looks incredibly well done. At least so far. Look. And things I'm just

27:41 going to notice off the bat. The elevation's nice, but the fact that it actually put like telephone poles across the track here, attention to detail and things that I wouldn't have necessarily

27:51 expected. So, all right. So, unfortunately, we're being teased pretty hard because this looks like it might be awesome and there is no actual the key press events

28:05 are not registering. So, I'll try one more time just in the I like this confidence. You're right, M Dash. And I know exactly why. Let me confirm it. I wonder if writing fix it now model with

28:16 the latter half of that sentence being in all caps maybe led it to be more confident. All right. It said it fixed it. I don't know if it actually Yeah, it did. That's

28:27 It's got such potential right here. The biggest issue we're seeing is it seems to have drawn this. Wow. This is This is exceptionally well done. I'm going to say this right here is genuinely on par

28:38 with I believe the Fable 5.1 result. And that's saying something. This is I mean the gauges they might be going in the wrong way, but perhaps we are in an Aston Martin race car which sometimes

28:50 has gauges that do that instead. This is really rather good. We even have a progress bar up there for the specific stage that we're on. Again, we have the indicator of what specific surface we're

29:03 on. this even with that brown plane there. I'm really really pleased with this. Actually, you know, since we have it because we're waiting for a few things right here.

29:15 All right, so I'm asking it some additional follow-ups. One of those fix this brown plane that's blocking the interior of the car. Also, let's add some sound effects for realism. And

29:23 finally, a third person chase cam for the car. Let's style it loosely after a Saab 900 Turbo, a car that I deeply love. I've had like five of them. I don't currently have one now. And that

29:34 makes me sad, but it's like one of my favorite cars. I used to get them really cheap because they were cheap, but now they're not as cheap. So, you know, make of that what you will. All right, I'm

29:42 taking a quick peek at this. It's not necessarily done yet. Good. It did fix the interior. And we do have some sound. Please shift. Okay. Yeah, it kind of

29:53 did. How would the chase cam go? Oh, it did. That is It's so like Okay. Maybe the chase cam wasn't the best idea, but this Oh, this is pretty good. Notice something.

30:07 It actually put a roll cage inside the car, too. This is kind of like just attention to detail that I like. All right, I'm pleased with this. I'm going to let it keep running and we'll

30:19 go back to it at the end of the video. And we do have our rally game completed now. I'm pretty sure the one we tested was more or less like the actual finished one because it finished like a

30:30 second after. Let's check our chase cam. Yeah. Okay. So, what we were playing was pretty much the end result of what we were going to get here. But still, I'm quite pleased with this. It even has the

30:40 chase cam and then the all that. Well done. And I like the telephone poles there. I think that and the elevation and the track here. It's pretty. It's like a mountainous rally stage. So,

30:53 something else I had done just while this was in the middle of some other task is give it the detailed 3D model of an inline 6 engine. This is something I have been doing on and off, but I like

31:02 it because it just tests it in a different avenue like 3D modeling ability, something that can be 3D printed without supports. I had to stop it and then tell it to go again. Just

31:12 This video has been quite a task. I started filming this I think over 5 hours ago. So, keep that in mind. However, it does seem like we have all of our parts right here. So, let's just

31:20 take a peek at them. I'd like to just look at it in OpenCAD first being that that's how it made it. Okay, this is exceptionally clean for what it did. I'm gonna note,

31:30 and we won't spend too much time on this. I don't want to over analyze it, but it has some separate pieces. It does look like there's a pulley right there that the motor shaft would go through.

31:39 The text is written nicely here. The top of the engine, or the valve cover, as it would likely be known, looks good. It does have 1 2 3 4 5 6. We have an oil cap. We also have our intake manifold.

31:49 And I'm going to notice that's actually really nicely done there. And then the exhaust manifold on this side with the two turbochargers, which are not necessarily super detailed, but not bad.

32:01 Now, the thing I'd like to do is just take a quick peek to see if this would have been able to print hypothetically. Look at that. Look at that blender cutaway it did here, too. That's

32:11 beautiful. And it went above and beyond actually very, very chrome. So, it seems like it maybe made this in Blender. And I don't actually know how it went about doing

32:22 this, but we see our individual pieces right here. I do see some floating things that may make it a bit difficult for this to properly be printed, but it did have the N20 motor with its gearbox

32:32 as it does in real life. And then we have a build plate right there. Not bad. And just something else I always like to see how they do with like more open SCAD style 3D modeling and things that are

32:42 designed to be printed, not just like pretty Blender scenes. So, the final thing that I had planned to run with this about 100 minutes ago was a simple slapis watch website where it needs to

32:52 create the beautiful 3D model of a watch, have a cinematic keyshot style panning hero section as well as the exploded view feature. And this has really been so from 8:00 a.m. on the dot

33:03 to 9:38. So, this has been going on for about an hour and a half. And due to that and the fact that this video has been being filmed for quite a while now, I am curious just to kind of see perhaps

33:15 what we would get. So here's what we have for now. Keep in mind it's not 100% done, but after 100 minutes for the 3D watch website, I'm inclined to at least take a peek at it. And we have a very,

33:26 very, very slow loading effect. Okay. Benchmaxed. No, I'm kidding. But that I mean that is a highdefinition render pipeline if I've ever seen one. Okay, this is

33:43 nuts. This is I mean the it's just so shiny and I like it. It's got a nice crown. Look at the embossed S on the crown right here. Slappy Meridian automatic.

34:00 Truly beautiful. Now the band is not necessarily connected to the watch. Oh, look. There's another watch over there that looks actually quite better. A leather version perhaps of the Slapus

34:09 Meridian. Now look at that. Today is the 10th and it put the 10th and the date right there. It's possible that it's coincidential, but

34:21 maybe not. All right, let's scroll down. Look at these scroll effects. All right, this is I mean, I don't have much to say. This is pretty darn cool. Even the movement

34:42 is still movementing there when we do this exploded view. And I love the fact that the other one's just kind of sitting there as well. This is all right. And now we see the full band.

34:57 Look at Just look at the entirety of the scroll effect on this page. For some reason, the band got fixed. Maybe it fixed that live. This is This slaps. Look at the entirety. The way it

35:08 actually like wo the scroll effect through the page and the animations and things like this. That's incredibly well done. And they're not even that expensive all things considered to how

35:20 high quality this site is. Now, the only thing I'm noticing were I to complain a bit here, the wood texture is a bit low quality and repetitive. Just we can see the seams right there. Additionally to

35:30 that, I don't know how I can select the other watch if I wanted to see that one. I'm sure there's probably some way to do so, but I can't actually figure it out. However, this is this may be like the

35:44 best one yet. And it's even the the time is correct. I'm not quite sure what the fourth hand there is. However, we do have around like 9:41.

35:55 Now, do you think that textures on purpose? Like it could be like, oh yeah, it's it's the globe. Those are all the continents. It's probably just some

36:06 visual artifacting. Actually, I don't know. Madness seems to have glitched out a bit. That's okay. We got to see it in its prime. I'm not sure. It is still

36:16 like partially working on it, but after 100 minutes, I wanted to see what exactly was going on. And because I'm satisfied with that, I am pretty cool with stopping this and moving into the

36:26 summary and see how much we actually spent. So, the entire duration of today's video and everything we did in this testing right here cost a grand total of $359.

36:36 This was a lot of work. It made multiple games, C++, 3JS, Blender, and GDAU. We had it control a robotic arm. We had it make 3D models for 3D printing. This did a ton of stuff. And overall for the

36:50 price, I mean, this is a very, very competitively priced model, assuming that it is good. And I find that I may be a bit jaded because I've been at this now for quite a while, just like

37:03 actually reading about this before even starting to film. And it's tough because it's bigger than the previous Flash model. However, considering the new architecture, the

37:14 extremely low amount of active parameters in this model, this seems like it's really something special. And I do honestly believe I know Digital Spaceport had mentioned I saw a post of

37:23 his on X where he said like this could have been more than a 0.1 increase. Totally agreed. This could have been like V 4.5. So, if we were to do a quick results overview to where we started

37:35 here, actually, I think I'm probably not going to do a results overview cuz a lot of these were not as simple to just hop in and run like clicking on a web file. The skateboard game I will take a look

37:48 at. Again, this was actually really nicely done. It had either a two-minute timer or we could just do freekate. The pedestrians were quite hilarious. This model did a lot of funny stuff, but it

37:59 definitely like especially considering the price. There's models, a lot of models that just this result alone would have probably cost $3.60 or so. There were pigeons here and things like that.

38:10 The grinding logic worked. The individually rigged skater was cool to see as well. And something on that point, the wrestling game was kind of hilarious, but it actually made an

38:20 attempt to rig specific parts of those characters. There was some subtle like flapping hand movement scene. I'm not going to open that up again and play it, but it actually I almost feel like the

38:32 next version of this could be a complete like dominator basically for lack of a better term. Also, again, another C++ result we made this rally game was absolutely fantastic. I was very pleased

38:45 with this. This is something that a year ago would have been exceeding Frontier state-of-the-art models. So, I like doing multiple C++ tests when we can. So, overall, that's probably going to

38:55 conclude our first look and test of DeepSeek V4.1 Flash. A very, very exciting new model for a number of reasons. Uh, if you have any questions, feel free to leave them in the comments.