Qwen3.8 27B + DeepSeek Harness BEATS Opus 4.6 in Claude Code

0:00A 27 billion parameter model running

0:02locally on a single GPU should not be

0:05able to compete with frontier level

0:07models like Claude Opus 4.6, but that is

0:11exactly what I've been seeing with the

0:13new Qwen 3.8 27B.

0:17Today, I'm going to be testing it at

0:19every single reasoning level, optimizing

0:21the model for maximum speed in Unslaved

0:24Studio Desktop, and then connecting the

0:26exact same local model into the brand

0:28new DeepSeek Harness for agentic coding.

0:32We'll start with testing the different

0:33reasoning levels, and then towards the

0:35end of the video, I'll be giving Qwen

0:37and Opus the same massive prompt to

0:40build a Minecraft clone as somehow this

0:44small model is beating Opus 4.6 and

0:48Claude code. Now, really quick, I don't

0:50normally like to go over benchmarks for

0:52too long in my videos, but this is an

0:55absolutely amazing website. If you've

0:57never checked this out before, this is

0:58artificialanalysis.ai,

1:01and they do an excellent job at

1:03comparing the different models in every

1:06single different model from every

1:08company. And down here, I just want to

1:11really quick show you guys, I added

1:13Claude Opus 4.6 with the max reasoning

1:16level, and it is actually scoring lower

1:20in the intelligence index [music] than

1:23Qwen 3.8 27B. And then coming down here,

1:26this is another artificial analysis

1:28index open weights, and same thing, it

1:32is scoring lower. If we go to this,

1:35about the same, and then text only,

1:37about the same. So, unbelievable that we

1:41have a 27 billion parameter model that

1:43we can run on local hardware, and

1:46generally affordable local hardware as

1:48well, is absolutely incredible that we

1:52have a 27 billion parameter model that

1:54is beating some models that are in the

1:57trillions of parameters. Absolutely

2:00insane. Now, moving over to one of the

2:02programs that we're going to be using

2:03today. This is

2:06Now, I've done videos in the past of how

2:08to install Unslaw Studio. Unslaw Studio

2:11also comes with Unslaw Desktop, but this

2:14is just the desktop app for Windows, and

2:16I like this app over LM Studio for

2:20specifically for Windows because it

2:22automatically optimizes the model to be

2:25almost the perfect weights and all of

2:27the different settings so that you can

2:29get the most out of the model and

2:31optimal speeds. Whereas in LM Studio, I

2:34was having issues with MTP generation

2:38and different speculative decoding

2:40issues as well in LM Studio. So, if

2:42you're unfamiliar with Unslaw Studio,

2:44check out my other video, but in this

2:46video, I'm just going to do a quick

2:47rundown of how to install and get Qwen

2:503.8 27B running. Now, my system, I have

2:54a 59 with 32 gigs of RAM. However, you

2:57should be able to actually run Qwen 3.8

3:00with a pretty decent context window with

3:03a 3090 or even a 4090 with 24 gigs of

3:08RAM, and there are even quants where you

3:09can run this at 16 gigs of VRAM if you

3:14have something like a 4080 or a 5080.

3:18So, if you go over here, you can go to

3:20this model hub, and you can go ahead and

3:22find Qwen 3.8 27B. Now, we're going to

3:25be using the Unslaw version because this

3:27is the best version that I personally

3:29know of. They are always releasing new

3:32updates trying to optimize their

3:34different quantization levels to make

3:36the model the best that it can possibly

3:39be. So, right here, I've actually

3:40downloaded a few different models.

3:41Today, I'm going to be using the 4-bit

3:44quant because I feel like that's what a

3:46lot of people are going to be able to

3:47run on their systems just to show you

3:48guys what this model can do even at a Q4

3:52level,

3:53>> [music]

3:53>> UD quantization. And UD stands for

3:56Unslaved Dynamic. So, essentially, they

3:58are taking some of the weights and

3:59trying to optimize it for the maximum

4:02potential. So, once you've got the model

4:04installed, you can go ahead and load it

4:07here. Now, let me just show you guys

4:08some of the settings. There are a few

4:10settings that I would recommend

4:11changing. So, on the right side here,

4:14context length. I'm using a pretty

4:16healthy 133,000

4:18context length that fits really nicely

4:21into my 5090s VRAM. Now, if you're

4:25running with 24, you could probably

4:27actually get away with this about the

4:29same context length. You might have to

4:31pull it down to like 100K, but that

4:33should still be very usable for what

4:35we're trying to achieve today. Now,

4:37right here, this is going to be key.

4:39This is the KV cache D type. This is

4:42quantizing the actual context length,

4:45and we are using a Q8 because that is

4:48where you [music] essentially have

4:49lossless quality. You go any lower than

4:53this, like Q4, your context window is

4:55going to start deteriorating just a

4:57little bit. However, Q4 is still very

4:59usable if you're trying to squeeze a

5:01little bit more context window out of

5:03your system. Now, a few other things to

5:05change. This should automatically use

5:08MTP with the auto setting. However, I

5:11would just set MTP, and then [music] the

5:13fastest I found so far with the draft

5:16tokens is two. That's just telling the

5:19MTP for speculative decoding how many it

5:22should use. And then parallel slots, I

5:24changed this to one. [music]

5:26And then the only other settings that I

5:29would recommend changing is down here.

5:31These are the different settings that

5:33Unslaved recommends for a thinking

5:36model, which is what I prefer to use.

5:38And then [music] this is going to be

5:40very important for the deep seek

5:42harness. I'm essentially just maxing out

5:45the max [music] tokens for the output

5:48length of every single prompt so that it

5:50doesn't run into any issues when we go

5:52over to deep seek harness a little bit

5:54later in the video. So, running a few

5:55tests, I just ran some tests in Unsloth

5:58Studio itself. I just wanted to see what

6:00it could do and I'm essentially making a

6:02Tetris clone. I'm giving it a pretty

6:04simple prompt, [music] make a Tetris

6:06clone, and I'm doing it with no

6:09reasoning, low reasoning, [music]

6:11medium reasoning, and then extra high.

6:13Those are the only reasoning levels that

6:15are available in Qwen 3.8 27B. Even

6:19though there is a high level, you set it

6:22to this high level, it's just going to

6:23default to extra high in the Unsloth

6:25Studio. You've just got to stick to low,

6:27medium, and extra high. So, this is the

6:29test that I ran with no reasoning

6:31involved. And all I did was I added this

6:34code, which is pretty cool. This is

6:36something that's included in the Unsloth

6:37desktop app. So, I am running this right

6:40now just directly in Unsloth. That is

6:43something you can do. You're not doing

6:45any crazy coding tasks or you just need

6:48a quick local model for some prompting,

6:51you can easily just use this desktop

6:53app. You don't have to use the deep seek

6:55harness that we're going to be setting

6:56up a little bit later in this video. So,

6:58this is pretty interesting. I ran a

7:00Tetris prompt. I just said, "Create a

7:02Tetris game in HTML."

7:04And I did it with no reasoning and it

7:06went through and made all of this. I had

7:09to try to ask it where the file was and

7:11it finally figured out how to get me the

7:12file. And here's some of the stats. This

7:14was running at an average of 100 tokens

7:16per second, which is

7:18unbelievable for speeds for this

7:20quantization level and that big of a

7:22context window. So, here's the Tetris

7:24game that it ended up making in HTML and

7:27this is actually really solid. It looks

7:30really good. It works very well. So,

7:32this is with no reasoning on and this

7:35might be what you want to default to

7:37when you're like essentially doing some

7:39quick coding is just turning off that

7:41reasoning. However, I do like to test

7:43these with the different reasoning

7:44levels just to see what it can make.

7:46This is honestly a pretty solid clone

7:49overall. Now, the one that is the best

7:51was actually probably the low quant.

7:53We're going to check that one out in a

7:54second here, but this is really, really

7:56solid for what it did. And it is

7:59honestly pretty interesting because this

8:01quant model seems to really take [music]

8:04the effort to a whole 'nother level.

8:07Like I've seen less effort sometimes

8:08from Opus, which is honestly pretty

8:10crazy than this small model. Like it'll

8:13make sure that it creates a really good

8:16generation. And as you can see, this

8:18honestly looks amazing with zero

8:21reasoning on. Now, moving on to the low

8:23reasoning level. This one was pretty

8:26interesting. It ended up using almost

8:29the entire context window, which is

8:33crazy, first of all.

8:35Like 124k context window.

8:38However, it made an absolutely

8:41incredible Tetris game. It went through

8:44and it took quite a while, but it was

8:47running at a really nice 110 tokens per

8:51second. And what it produced honestly

8:53blew me away. Now, this is actually

8:54nice. It gave me an HTML preview here.

8:58So, I can just expand this and play it

9:00right here, right in Unslaw Studio. This

9:03one has the best graphics I've seen. It

9:06looks better than the last generation.

9:08And I think it even had a full animation

9:12when the block or when the line breaks.

9:14Let's go look here.

9:16>> [music]

9:16>> So, let's get that line.

9:20Yeah, see that animation right there?

9:23So, this one it spent like quite a long

9:25time to build this. Like it probably

9:28took 30 minutes, but it put every single

9:32ounce of effort that it possibly could

9:35into making this [music] Tetris clone.

9:37And this might be one of the best Tetris

9:40clones that I have [music] ever seen

9:42with a local model. In fact, this

9:44probably is the best Tetris clone I've

9:46ever seen using a local model. Like even

9:49better than some generations from Opus,

9:51which is

9:52actually crazy. Now, moving on to the

9:55medium reasoning level, this one was

9:58very interesting to me.

10:00It only used 10K of the context, and it

10:03practically built this entire game in

10:06less than 2 minutes.

10:08And also, the token speed was

10:11unbelievably fast. 141.8

10:15tokens. That's the highest I've seen so

10:17far in any generation I've tested. Yeah,

10:19as you can see on the right here, it

10:20only took about 39 seconds to create

10:23this generation. So, I'm not sure if

10:25there's something wrong with the

10:26reasoning level with this specific

10:28version of the Unsloth

10:31when 3.8 27b. I know we're on version 3

10:34now, but it did actually end up making a

10:37Tetris game here. And honestly, this one

10:40isn't too bad. This is with the medium

10:42reasoning, and it did the fastest

10:44generation, and it still looks really

10:46solid. And it works just as well. Let's

10:50see if we can get a line to

10:53go here.

10:56And yep, line disappears. It doesn't

10:58have any nice graphics or anything, but

11:01it still has the preview block down at

11:02the bottom. It has the down, it has the

11:05drop function, it has pretty much

11:07everything a basic Tetris game would

11:09have. And it's kind of actually crazy

11:11that it generated this this quickly

11:13using the medium reasoning level, where

11:15I would have expected the low to be much

11:19faster. So, there's something very

11:21interesting with the medium versus low.

11:24Low, for some reason, seems to be higher

11:27reasoning level than medium, for

11:29whatever reason. I'm not really sure why

11:31that is. And finally, we have the extra

11:33high reasoning level. Now, this one used

11:3560k of the context window and it created

11:40a pretty decent Tetris game. However,

11:43again, it's very interesting that the

11:44low setting actually used more context

11:47and reasoned for longer than the extra

11:50high. Here we can see we got a decent

11:5299.2 tokens per second. All right, and

11:56here's the generation for the X high.

11:59And this one is interesting cuz it looks

12:00very similar to the medium generation.

12:02However, it took much longer to create.

12:06It looks almost identical to the medium

12:09generation. And one thing that I noticed

12:12is I cannot spin the blocks more than

12:14one time. Now, let's see if it'll let me

12:16spin it with the Z and

12:19the X. Nope. Looks like you can only

12:22spin the block one time. So, even this

12:24one on the high reasoning level had all

12:27kinds of issues. So, so far, it looks

12:31like low is the winner. If you want the

12:33fastest [music] generation possible, you

12:35got to go with medium. And then the no

12:38reasoning one was also pretty decent,

12:40but I'm still going to give low the win

12:42here when it comes to the reasoning

12:44levels. All right, so next thing is

12:46hooking this into Deep Seek Harness. And

12:49this is pretty awesome. So, essentially,

12:51what you've got here, you can go ahead

12:53and install this pretty easily if you go

12:56to their website and follow their

12:57GitHub, and you can install this on

12:59Windows and it uses essentially Node in

13:02order to install this Deep Seek Harness,

13:04and you can use it in Windows or Mac.

13:07So, if you go to this models tab here,

13:09you can go ahead and set up your Unsloth

13:12Studio to connect to the Unsloth Studio

13:15using your configured model. So, in my

13:17case, I have configured an API key in

13:21the Unsloth Studio. I'll just show you

13:23guys this real quick. You can go ahead

13:25and go to Unsloth Studio, and go to the

13:28API section. Go ahead and type in like

13:31Deepseek [music] Harness, for example.

13:34I'll just say Deepseek Harness 2 cuz

13:35I've already set this up.

13:37>> [music]

13:37>> Create a token. Generate that here.

13:40You'll copy this token, and you'll copy

13:42that here into the API key, and then you

13:45can set it up so it's fully locally.

13:48Just make sure you point it to this

13:49address. By default, Unsloth will use

13:52port 888 at v1, and then OpenAI chat

13:57completions, and then I have simply just

13:59set up this model here.

14:03Now, unfortunately, right now, this

14:04isn't a developer preview, so getting to

14:06some of the settings is a little tricky.

14:09So, you are going to have to type in a

14:11specific command in your PowerShell

14:13window in order to set up the different

14:15reasoning levels. And that is this

14:17command here, the notepad

14:19env.userprofile.dnsah.settings.yaml.

14:23And if we go to this, just really quick

14:25show you guys, these are the settings

14:27that you want to use with the Qwen 3.8

14:3027B model. It essentially tells it,

14:32"Hey, I want you to use these reasoning

14:35effort levels, low, medium, and extra

14:37high," which is the ones that are

14:38embedded in the Qwen 3.8 27B. I'm giving

14:41it my max context window, and then max

14:44tokens, and then also, you want to make

14:46sure you use this input here and do text

14:49and image, otherwise, it's not going to

14:51be able to use the vision portion of the

14:53Qwen 3.8 27B. Now, I'll put those two

14:57commands in the description below so you

14:58guys can just copy all the settings and

15:01get to that settings menu pretty

15:02quickly. Now, moving on to this Deepseek

15:05Harness, this is a really cool program.

15:07A lot of people say it's kind of like

15:09Claude Code, but I think it's more like

15:11Codex, where you can have a workspace

15:13area or a project, and then you have

15:16chats that you can have underneath it,

15:18and you can start new chats underneath

15:20that, just like Codex. So, right here I

15:23have a couple examples. I did another

15:25Tetris test using the Deep Seek harness

15:27just to test it out, and I did it with

15:29the low-effort level. It went through

15:31and created the full Tetris game. Now,

15:34there were a few problems, like it was

15:36cutting off at the top, and I had it

15:37fixed that. And then I had to

15:39essentially recenter it down here, and

15:42it did create a pretty nice-looking game

15:45here. Let's just do a quick test. This

15:47is your Tetris clone here, and it looks

15:49pretty good. Pretty solid. Everything

15:51works good. It actually has nice line

15:54animations, as you can see there.

15:58And everything works, and yep.

16:01So, pretty solid, and it's fully

16:03centered on the screen. Now, that was a

16:05problem, it was like centered over here.

16:07So, that is the one thing I have noticed

16:09with the Deep Seek harness. You do kind

16:10of have to go back and forth with it

16:12just a little bit, but that's similar to

16:14like things like Codex or Claude Code,

16:16where you just have to go back and forth

16:18with it until it to fix certain things.

16:20But, to set this up is actually pretty

16:22cool. You can say add workspace here,

16:24and then down at the bottom, it's

16:25actually going to load a node, so node

16:28program, and then you have to go in it

16:30and define what your folder is, just

16:32like Claude Code, just like Codex. So, I

16:35could do like uh just test folder here,

16:38and then right now we have a new

16:40workspace, and I could tell it something

16:42like I could do just do a quick example,

16:44create a notepad app. And it's going to

16:46go through, and it's going to connect to

16:50>> [music]

16:50>> our Unsloth Studio here, which you can

16:53even open up this full API window, fully

16:56monitor what it's doing, all of its tool

16:58calls and everything coming from the

17:00Unsloth Studio. And then if we go back

17:02to the Deep Seek harness, we can see

17:05it's going through the full thinking

17:06process, trying to decide what it wants

17:09to do first, thinking about the empty

17:11directory that I gave it. It's a pretty

17:14broad prompt, so I'm actually kind of

17:16curious to see what this will do. All

17:18right, I didn't that really quickly. I

17:20guess I didn't give it a whole lot of

17:22instructions, but it built this notepad

17:24HTML. Let's go ahead and see what this

17:27built for us here. All right, so it just

17:29built a pretty basic notes app, which is

17:31interesting, kind of cool. Untitled

17:34note, let's just call it notes test. I

17:36don't really expect this to work very

17:38well cuz we didn't really give it a

17:40whole lot to work with here. This is

17:41pretty cool. It literally built this in

17:43a matter of like 30 seconds and deep

17:45seek harness. So, that was just an

17:47example I was going to show you guys on

17:48how to get your different projects set

17:51up. However, with the right prompting

17:53and the correct harness, you can build

17:55some pretty cool things. So, let me just

17:58show you guys what I built here. This is

18:00one of my prompts that I've been working

18:02on for a very long time. I'll provide

18:05this prompt in a link in the

18:06description. If you guys are familiar

18:07with my videos, I like to share all my

18:09prompts with you guys, but this is the

18:11full Minecraft prompt that I have been

18:14slowly developing and making better and

18:17better for the sake of doing my testing.

18:19And what it built is honestly pretty

18:22crazy. So, the first run-through, it

18:25actually built a [music] Minecraft clone

18:28that was playable and working. Like, I

18:31wasn't even expecting it to be fully

18:33playable at all, and it actually was.

18:36Now, I have kind of been messing around

18:37with this. I've been going back and

18:39forth with it, back and forth with

18:41different things like like making the

18:43animals look accurate, different things

18:45like that, and it's gone through and

18:47it's fixed pretty much everything that

18:49I've asked it to do. And I haven't made

18:51too many changes here. Essentially, just

18:53changed like the night cycle and pretty

18:56much the animals, and I made the zombies

18:59and everything more like realistic and

19:01functional, but that's about all the

19:03different prompts that I've changed so

19:05far with this whole thing. And it

19:06created this file that is just an HTML

19:09file that that can run in our browser

19:12here. And honestly, this blew me away. I

19:16was running this and I was like, "Wow,

19:19this [music] is absolutely crazy." Like

19:21it's got different like crevices, like

19:23cave system.

19:25It's even got a full crafting system.

19:27[music]

19:30And it was running at a really decent

19:33speed, too. Like about like 120

19:37tokens per second or so. It's even got

19:39like big mountains, nice rivers, the

19:43like water works where you can like go

19:46up and down the water. It's even got

19:48like a hold so you can see like the

19:49block that you're holding. The one

19:51thing, there's a few bugs like right

19:53here, I can't jump out of the water if I

19:55don't put a block there and the water

19:57doesn't fill in, but that is pretty

19:58normal for these generations. So, this

20:01is on par with one of the generations

20:04that I've built with Fable and also Opus

20:075, which I just thought was absolutely

20:10crazy. And like check this out. It's got

20:12like a full crafting system. So, if we

20:14go find like a tree here.

20:18The only thing that I think needs a

20:20little fixing with this is it takes a

20:23very long time to break the wood blocks.

20:27So, I'm going to speed this up and just

20:29come back so we can show you guys the

20:30full crafting.

20:33All right, so back in the game here. If

20:34I grab my wood blocks and place them

20:36here, it's actually going to create the

20:39wood. And then we do four here and it's

20:43literally created a crafting bench,

20:45which just blew me absolutely away. And

20:49I can set it, too. So, I can set it

20:51here. And then I can go into the actual

20:54crafting and create different tools like

20:57sticks. So, it's got like a full

20:59crafting system. Let's see if it'll

21:01actually even make the pickaxe here.

21:05Let's see if I can remember this. Yep.

21:07Oh, wow. It created the pickaxe and now

21:09we have a pickaxe and you can actually

21:11like see it in my hand. I mean, it's not

21:13the best animation in the world, but it

21:16does work as a pickaxe and you can see

21:18that the blocks are now

21:20racking and breaking way faster. So, it

21:23is working, which is

21:25actually crazy that I have a working

21:28tool system and that this bully

21:31generated that, too, by the way. And

21:34look how great this looks, too. Like, it

21:36did a really good job and it's

21:38essentially an endless world full

21:40generation and it even has a full night

21:44cycle and zombies and everything that

21:46come out at night.

21:49So, pretty cool. And let me just speed

21:51it up so I can show you guys the night

21:53cycle, as well, here. All right, so the

21:55sun's starting to go down and starting

21:56to be night time here and you can see

22:00that there are now zombies that have

22:02essentially spawned.

22:05And wow.

22:07>> [laughter]

22:08>> Okay. Well, we got them. But, yeah, it's

22:11got actual zombies that are essentially

22:14going to hurt you. It even has the

22:15creepers. Uh-oh.

22:17Just fell into like a cavern here.

22:19That's no good. Oh, there's the

22:21explosion. Creeper. So, it's got like

22:24explosions and everything. You can see

22:26the block animations, too. Like, some

22:29generations don't even create this like

22:31floating block animation when it like

22:34breaks or breaks a block, I should say.

22:38And yeah, so right now we're in pretty

22:40hostile territory here. There is quite a

22:43few. That's one thing I noticed with

22:45this generation. It created quite a few

22:47different zombies, but just the fact

22:50that it is a

22:51almost [music] nearly playable Minecraft

22:54game

22:55with some tweaks just honestly blew my

22:58mind. Like, this is just so cool that a

23:01local model is able to create this kind

23:05of generation that's something like even

23:07Fable 5 is creating something similar in

23:11this regard. Like pretty crazy. Let's

23:14see if that'll get me.

23:15Oh, almost.

23:17>> [laughter]

23:17>> Wow. So anyway, let's go ahead and check

23:20out now the Opus generation, which I was

23:23not impressed with at all. I don't know

23:26if they have [music]

23:27essentially nerfed the Opus one, but it

23:31was pretty bad. So let's go check that

23:33out real quick here.

23:35So this is the Opus generation here, and

23:38this is honestly one of the worst

23:41Minecraft experiences, worst Minecraft

23:44clones I have ever seen in my entire

23:47life. Now, to be fair, I did use the

23:50exact same prompt, and I did use Opus

23:53with low reasoning. However, I was just

23:55trying to match what Quen 3.827b

23:59would also be using. So I used both of

24:01them with low reasoning effort. And I

24:04don't even know what this is. Like this

24:06is honestly like one of the worst

24:10Minecraft clones, probably actually the

24:12worst Minecraft clone I've ever seen.

24:14Like I'm not even sure what I'm looking

24:16at here. So I went back into the chat. I

24:18was trying to give it the benefit of the

24:19doubt, and I said, "Okay, it finally

24:22loaded. However, this is one of the

24:24worst Minecraft clones I've ever seen."

24:27And it actually came back, and it tried

24:30to fix it. It actually did a decent job.

24:32It went through and thought for like

24:33about 8 minutes. But the first one that

24:36it did with that massive prompt, it only

24:39thought for like 3 minutes.

24:40>> [music]

24:41>> So I guess low reasoning effort is

24:43actually very low when it comes to Opus.

24:46So this is what the revised version

24:48[music] created for me.

24:50Now, to be fair, I will say the graphics

24:53look pretty solid.

24:54Like the graphics are not too bad. You

24:56can actually jump out of the like water

24:59area. The graphics look really good.

25:01However, there is quite a few problems

25:04with this generation. So, first of all,

25:07I was testing out the crafting system.

25:09So, and also, okay, this is the first

25:11thing I noticed. Look at these blocks

25:13that are floating.

25:15Like what even is that?

25:17That looks awful. I don't even know what

25:18that is.

25:19>> [music]

25:19>> So, I was trying to get the crafting to

25:22work. And I couldn't even get it to

25:24work. So, I got, you know, the first

25:26wood blocks, went to my crafting window.

25:29First of all, I was trying to figure out

25:30where to place these. I don't know know

25:32where it even went.

25:34>> [music]

25:34>> So, it just went to storage here, and

25:37then I tried to go and do this.

25:40And so far,

25:43it is not even understanding to make

25:47wood. Oh, there we go. It actually did

25:49work to make wood. However, I can't like

25:51grab it. So, let's see if I go out of

25:53this menu here.

25:55And yeah, like [music] I can't even

25:58pull it out of the crafting inventory.

26:00So, regardless of this really funky

26:03animation, it still does not work. I

26:07cannot get this to actually generate or

26:10pull

26:12from that storage. So, the storage and

26:14the crafting essentially don't work at

26:16all. I've been trying to click on this.

26:18I'm left clicking on it. Like I don't

26:20know how to grab these actual wood

26:23blocks. So, the crafting system is

26:25broken. Um this is another thing I

26:27noticed is when I hit the escape button,

26:30it brings up that menu. And then, this

26:33is one of the worst things with this

26:35that I noticed is the water. Let me see

26:37if I can find some better area before it

26:39gets too [music] dark here.

26:41And we get like some zombies coming out.

26:43Also, the night cycle is like really

26:45fast. Okay, here we go. This is a good

26:47example. If I jump into this water, you

26:50can see I'm not floating. Like it shows

26:53water on top, but I can't float up.

26:56See, I'm just jumping underwater. So,

26:58there's no water physics.

27:01So, that's another thing.

27:03And honestly, like I [music] said, the

27:05graphics do look pretty good. Like,

27:08arguably better than the [music] Quen

27:113.8 27b, but honestly, I'm giving Quen

27:153.8 27b the win here because it is

27:18actually more functional Minecraft than

27:21this. Like, I don't even know what this

27:22is. What kind of zombie is this? If you

27:24look at the side, well, it's kind of

27:25hard to see, but it's a four-legged like

27:28like horse zombie.

27:30Like, even this doesn't even look

27:32accurate to like the original.

27:34If I can actually get it there, there we

27:35go.

27:37And yeah, also, you notice that they are

27:39able to walk underwater just like me.

27:42So, and then trying to get out of here

27:44is a pain because there's no way to

27:47float. You got to like kind of make your

27:50way up.

27:51And yeah, pretty much overall, not very

27:54impressed with the Opus 4.6 generation.

27:58Now, it did do this much faster, but it

28:00is on, you know, Frontier hardware. We

28:03are running this on a 5090 with Quen 3.8

28:0627b. So, overall,

28:09I would say, like I said, Quen 3.8 27b

28:13wins in this test, which is absolutely

28:16crazy that a small local model is

28:20beating Frontier level models from

28:23Anthropic as well as OpenAI.

28:26So, anyway, that'll do it for this

28:28video. If you guys have any questions,

28:30let me know down below in the comments.

28:31I plan on going over some more local AI

28:34content. I'm also got a video planned on

28:37using GrokBot because I just stumbled

28:40across that today, and I think it looks

28:42really cool. It has a lot of really cool

28:45potential. So, stick around, stay tuned

28:47for that video. Subscribe to the channel

28:49if you found this helpful, and as

28:51always, have a wonderful day.