Ling 3.1 Flash First Test – Is THIS Really a DeepSeek Competitor?

0:00 One, two, three, four, six, seven, eight, nine, 10. The way it's like, however, today we're going to be taking a quick look at an interesting new flash model from Ant Groupoup, and

0:12 that is Ling 3.1 Flash. Now, I've tested a few of their models in the past, and one of them still sticks out to me as one of the more impressive ones I've seen, and we'll probably quickly just

0:22 touch upon that before we get into it. However, the main event for today is this 560 billion parameter mixture of experts model with 25 billion active. Now, this is not currently open weights.

0:34 However, they do mention they plan to open source or open weights it soon. And because they do have a very active hugging face right here with all of their previous model releases available

0:44 for download, I have no doubts that this will also make its way up here and join the rest of these. Now something interesting and it kind of tracks with DeepSeek V4.1 flash is this flash model

0:55 is significantly larger in size than its predecessor. So we can look right here and see 3.0 flash which if we click on that that's like a finance specific one. If we go right here this was a 124

1:07 billion parameter mixture of experts model and now the next flash version is 560 billion parameters and we kind of saw that growth as well with Deepseek v4.1 flash. So, it's cool because

1:18 they're more capable. However, the downside of that is it's less able to be run on some enthusiast hobbyist level setups now. So, before we get into it, do feel free to subscribe so we can get

1:28 that 100K plaque. And let's take a look at the first thing listed right here, which is of course a benchmark JPEG. Now, this is being compared against some outdated models now, at least in terms

1:37 of these two right here. But the rest of these are more or less current, and that's nice to see. I think Deepseek V4.1 Flash is probably the most fair competitor against this model that I see

1:47 in this chart. Again, I don't really pay too much mind to these benchmarks, especially since the Muse 1.3 result on Deepsw SWE. So, basically, I just show these because it's interesting to see

1:58 them, but with that, I'm more focused on seeing how it actually performs in some of our tests. Additionally, they talk about some of the demonstrations of work that it's done, talking about this

2:07 17-hour process where it built a Lua to x86 compiler from scratch and passed 97.8% of the tests. But it is cool to see more benchmarks now showing models performing over many hours to perform a

2:19 task and not just like simpler things cuz it shows a progression in wanting these things to perform autonomously and perhaps with less human intervention needed while they work. Then they do

2:29 also of course showcase some 3JS chops right here. So we will of course be testing this. I like to see that engine one right there. I'm quite curious about some of these look pretty good actually.

2:38 So I'll be excited to see what our subway FPS result ends up looking like. Ah the one thing I forgot to mention the model that I had tested from them previously that still sits with me is

2:47 this Lang 3.0 Tiny. This was an 8 billion parameter mixture of experts model with 1.3 billion active. This performed very very well considering its footprint. Additionally to that, there's

3:00 really not a lot of models in this space like this size as ane. I believe IBM has a granite 4 model that may be ane around this size and I think liquid AI has some as well. But this is kind of like a rare

3:13 size and the capability from this model especially considering the like it's ane. It's not like a dense 8B. So, this was pretty cool and I've still continuously seen decent comments about

3:25 this from the community like a month or two beyond its release date. So, this is an interesting model because it sits in a far less saturated size area. So, first up, let's take a look at our

3:35 browser OS. I ran this a while ago and I never actually looked at the result. So, okay, first and foremost, I have to say I'm pleasantly surprised with what I'm seeing. I do need to continuously remind

3:46 myself this is a pretty large model, so we should see some decent performance. First things first, it does have the correct date and time in the bottom right. Is there a right click? Oh, there

3:56 isn't. All right, we do have a start menu. Grand Theft Browser, very creatively named. Okay. Oh, it mpark 1.3 does. Okay, so check this out. In the background, I actually see

4:12 something. So, it almost seems like this mini map is covering the entirety of the screen, making it look like this was the game it made when in reality it's just a UI glitch. Now, maybe we will try to

4:23 have it fix that, but keep in mind I don't actually have that specific session open and it I don't know that I can resume it. So, okay, Star Strike was not all right. It's possible the app

4:35 icons don't work, but fortunately, we have our search f function feature. I was trying to say function and feature at the same time. All right, this is kind of cool. It's giving me

4:47 like paleibious vibes with these shapes. And it's weird because it looks like the background is just leaking through into this. So, which is fine. It It's It's a level of immersion. Is that an enemy

4:58 ship shooting at me? Yeah, it is. Blasted. All right. Not bad. Next up, male wave, which we have to manually type. All right. Simple.

5:11 As always, meta references to the apps on this specific operating system. Issue number 241. It gives us a little info about the wallpapers. And we have a little notification in the bottom right.

5:24 And it goes in our sent mail. The heck? Oh, we got a new we got a new email. Okay, that's good responsivity to our fake sent email. All right. Simple file browser. Can we open any of these? It's

5:41 saying procedurally rendered and I think it's supposed to show them here. And these just look like maybe the basis for our moving backgrounds terminal. Huh. Interesting. I typed it in all caps

5:55 and it still registered it. That doesn't always happen. Some most often they're case sensitive. So I like seeing that. Oh, that actually worked. Okay. All right. So this is our special

6:07 feature eightball. Press control K and type anything. That's actually kind of cool interesting and and a pleasant surprise in the interactivity. Can we close this with control K? We

6:19 can. Notes calculator. The nemesis. Some of the UI here is not like the best. 25* 41. That should technically be really it should be,25. Yeah. All right. Good. He's on

6:35 clock. Nice. And we even have the date. 1 2 3 4 6 7 8 9 10 11 one. System monitor. Not bad. And it shows us

6:49 our processes. Processes with the PIDs as well. Personalize. Oh, wow. Okay. Shuffle wallpaper seat. I think something is not quite working here cuz there look at the look at the similarity

7:03 here. Oh, hold on. Look at the similarity here. Oh, looked more similar when I closed out of this. Accent colors, sound effects. Not bad. Overall, this was acceptable. Um,

7:16 the biggest issue was the Grand Theft browser was just having some problems. So, let's look at some other results, but we'll probably have that fixed. The next thing that I had done was a C++

7:27 skate game. This initially, it did work. It compiled without any problems. It was just like the camera was halfway into the ground and it was not following the player. So, I yelled at it and said

7:37 like, "Hey, this is terrible." And then it fixed it. Something I will say is it fixed the issues very, very quickly when I had described them following making the initial result. Now, of course, this

7:48 right here is definitely not in line with what we've been seeing recently with a lot of the Frontier models like Claude Sonnet 55 that have been coming out. However,

7:59 it did like um it compiled without errors. It definitely is something that could be iterated on. And I mean still I think the like don't

8:13 get me wrong, yeah, this is bad, but the takeaway is like a year ago this would have been like they're evolving. So, and we have some like Yeah, that's always small model smell or just not good when

8:24 they get stuck at the end of the rail there and like just go up and down. We also bail a lot. Let's try to shove it and then the um the people are they're drawn on the wrong plane. All

8:44 right. So, yeah, that was that was a skate result. Now, everything else I've not seen. The Skate one is the only one I've seen so far. I didn't see the browser OS. And then everything else I

8:56 believe I ran is still open here and finished. I have not looked at any of these. So, let's do what is this one? Let's do a simple one. Now, this is the beautiful watch

9:07 website. We can see right here how many tokens were used for this. Let's take a peek at our Slapis watch website. As I said, this is probably going to be a quicker and more casual video, but I

9:17 hadn't tested anything in a while, and I had wanted to test this. So, it's always fun to test models. I can't tell if I want to say benchmaxed or not cuz I see things here that are

9:29 like not right, like the kink kink like leather band there, and then the way it's like, however, there are also some pretty good things here, like the crown is wellformed. The face is nicely

9:43 formed. The numerals are not in that stupid like crooked but symmetrical way that we sometimes see with less capable models. And additionally to that, the hands move well. And it does have a very

9:53 nice date selector. It's also placed on this. For some reason, models assume like this is what a high-end table looks like. You might have seen tables like these in like hotels and stuff. They

10:03 have I don't actually someone could please feel free to enlighten me as to like what the significance of this style with like elegance is I digress. Quiet obsession with light. The rest of

10:15 this is the black and white and gold aesthetic that we're used to seeing. Okay, we have the Aurelia for 48.50 and the Meridian for 29.50. So perhaps a cheaper

10:29 brand. Um, okay. It doesn't seem like we can actually interact with these in any larger view, but they do both render and

10:39 look acceptable. Now, did it do the Yeah, it did. It did do the exploded view. And I have to say, this is definitely a respectable result for a model that will be open.

10:53 And look at that. Though it's hard to see, it did write slapis on the rotor there. It's just unfortunately the material is not rendering from that angle. and then a simple footer. The

11:01 next test I gave this was a more difficult one, a blender and gdau test that we've only started doing very recently. It's the one where it needs to make a pool party game where you jump

11:11 off a diving board and then land in the water set in the backyard of a house and it needs to have emphasis on like water effects and make everything look good. I'm just going to ask it run it for me.

11:21 I had to stop this a couple of times cuz I was doing something else on the system and it froze in its entirety. So, I had to hard shut it down and then open this back up and say continue. And it seems

11:31 like it did. So, we'll see. It's not bad. It's definitely not rendering the entirety of the UI properly. I also don't hear any sound and the speaker is on

11:45 and not muted. So, I'm just going to press enter. Okay, that unfortunately did nothing. I'm going to try to click Oh. All right. So, enter did nothing. Space was the choose your cannonballer.

12:03 Well, I can't see anything. So, let's play a sub game from within this game called Let's try to guess what is missing from the scene here. um properly done humans. However, we can assess that

12:28 we have three individual judges actually showing the this thing has a pretty big problem with making human models just based off of what we've seen with this as well as the

12:43 skate game. They basically had the same form of like horrific geometry ascribed to them. However, this is a it's an acceptable result if we ignore the humans. Like the grill has

12:56 smoke coming out of it. It has a nice like light pastel color palette. Round one complete. There is game logic. Okay. I mean, I'm going to just try to Oh, cool. Okay.

13:11 The camera shake is is quite aggressive. That's going to be a higher score. Yeah, I would even have to agree with the judges on that one. All right, this will be our last one.

13:31 That might have beat the previous one. Six. Five. I think that was the same thing. Okay. Overall, it it had good elements and it had bad elements. All right. I've given it some

13:44 feedback there. We'll see if we get a improved result. All right. I'm happy to report that in a little under 14 minutes, it did make the proper needed changes where now we see these are

13:55 actually humanoids. Also, there's a little dog here wagging its tail. I don't know if that was present before, but if it was, it may have also had some of the same geometric issues. So, let's

14:06 uh All right. There's still the UI is a bit buggy, but that's okay. All right, let's see. Good. And they're walking, albeit basic. So, oh, cool. Look, then the judges get

14:21 excited and the other players are the ones scoring. That makes a bit more sense. All right. I don't know who I chose, though. So,

14:30 cuz I didn't really do anything there. So, and the dog is visible as well. This is It's simple, but it's this is not bad. And part of the reason I like doing this is because it's new

14:43 enough that this would not have been in the training data for this model. So there's no real benchmaxing capability for this. And the water is not bad. There's like a floating beach ball in

14:55 there or something. And it's actually moving with the water current. So overall, I'm actually I think this is us. We'll do one more and just what? Okay,

15:07 I'll take it. I'm actually happy with it. It seems to quickly have the ability to fix things. The Subway FPS game. Now, based off of what we've been seeing um with some of the humanoid models, I am a

15:19 bit concerned that the zombieoid models here are going to be quite bad, but the rest of it might be good. So, let's check the Subway FPS. Okay, deadline. So,

15:40 the good thing is it seems to have understood what humanoids should look like with this specific result. The bad thing is this is a pretty poor result based off

15:52 of what we've seen from competing models. So, this one not so much. You can right click to zoom in and aim. And let's see how the train logic does. So,

16:06 All right. Now they should get out of the Did the train just At least it didn't back out of the station.

16:26 Okay. Well, said crumpled trash. All right. In 28 minutes or so, we have a supposedly redone Subway FPS game. And that was the one that was really like

16:40 kind of disappointing. So, I'll be excited to see if it made it a lot better. Definite improvement. The models are still relatively simple.

16:56 Unfortunately, it's lagging when we shoot, but like look at the tile work and things like that. The weapon models are better. The big problem is the lag here, which

17:07 is kind of a shame. I don't know if I told it there should be bullet holes in the environment. I that's that's a shame because I think the takeaway here is it did definitely

17:21 show ability to improve results that were kind of not very good. We have a reload animation. I just want to see if there are bullet holes. No, I don't think so. That's okay. Well,

17:34 it's not okay, but okay. Still the floor is bleeding when we shoot it, which is absolutely horrifying. Okay. And our train is a bit more

17:56 detailed as well. Let's see if that weird gl that weird glitch happened where it like moved. Okay. No. Yeah. All right. So,

18:12 unfortunately, not quite, but it was a significant improvement over the first result. Let's try to see if it can fix the issue we noticed with the GTA game in the browser

18:25 OS where the map was covering the entirety of it. So, we couldn't really actually see the game. And this is something where it's not continuing from when it had initially created this. So,

18:34 it's going to have to just read the script and figure out what the issue is. So, I had mistakenly run that session from the Subway FPS folder, and it was very confused. It was like, I have no

18:44 idea what you're talking about. There's no mini map. There's no GTA game, and I was like, "Yeah, it's in there. Single file HTML." And I was wrong. So, we're going to just go to the correct folder.

18:54 All right. Let's see if now without me making a mistake of what file I'm pointing it at, it can fix this. So, in 1 and 1/2 minutes, apparently, it's fixed. Now I need to open it from this

19:03 folder because I moved the location from the one that was open there. Every time I restart this system, the chat GPT application resets itself as the default app for HTML files, which is to be

19:15 honest with you just infuriating. So little note there of rage. Now let's see. significantly. Okay, significantly may have been an

19:30 overstatement. It is reminiscent of perhaps a movie in which one is not of right mind. Um,

19:47 now this is obviously something's quite wrong, but it's also in a way kind of cool like as an effect that's just like this would be like an effect someone made and you'd be like,

20:00 "Wow, that person's creative and skilled like to have thought of such an odd thing." I'm almost happy it it's this weird because

20:13 we're definitely getting a unique take on this. I told it that it looks like a TV movie depiction of what being in a black hole would look like and hopefully that'll help it fix the problem.

20:27 Okay. Um, better. Still not great. Okay. We had a traffic accident which we're actually able to see. So, it did implement that like if you hit a car, you get Did it put mesh colliders on the

20:40 buildings? It did. Okay. So, let's just see if we can interact with one of these pedestrians. Oh, I think we hit a car. Let's just Yep. Okay. Well,

20:57 interesting. I don't know if we'll actually be able to finish this without it maybe timing out or having some errors, but I'm going to attempt to give it the complete full

21:08 Street Yeet build prompt. This is one where it needs to use Blender and GDAU, making it a bit more potent. Hopefully, we'll see how it does. And this is one that's just a bit more defined in terms

21:18 of like an agentic prompt. All right. So, our last test has completed and I noticed because it one thing I can tell you without having seen

21:32 this yet is that the audio is quite loud. Oh, okay. So, it seems there's a bit of backwardsness here, but that's okay. And you'll hear in a second. LET'S JUST

21:43 YEAH, it is quite loud aside from the inverted movement. Oh, and text. Okay, good. There are mesh colliders with the cars and the building. Something I'm noticing

21:55 here just out the out the gate. It does seem to have some like spatial issues where things are backwards here or mirrored or inverted. And that's something I think was a fairly common

22:06 theme across a number of different tests with this. However, overall, this is a respectable result. If the wheels on the vehicles were actually arranged the right way and if our player was moving

22:16 the right way, it definitely would have improved the overall experience here. I guess let's just But it is good that they're actually mesh colliders so like people just

22:30 bounce off instead of go through. That's a nice Oh, cool. I don't like the lighting, but you know,

22:45 can we d W? Oh, yeah, we can. Wow. Oh, okay. And then the draw distance uh captured that individual from our view. The fire HYDRANT GOES

23:01 the bench. I don't think we can. Yeah. Okay. There's only been one time we could ye a building, and I'm pretty sure that was because I'd put it in the prompt.

23:10 All right. Overall, I think that's probably going to conclude what's inevitably going to be a lighter test than normal on Lang 3.1 Flash. I wanted to test this because it's a new model

23:20 that will eventually be open weight and there hasn't been a lot of openweight model releases lately, at least of note. So, I wanted to get something out there and I'm interested to see if this

23:30 improves prior to the openweight release. It seems currently were I to give like a bit of a results overview. For the size, I'm not blown away with the results that it's giving us. It

23:41 seems almost undercooked a bit. It's possible this is maybe an earlier preview to gather feedback or see as often is done. So overall, I don't have a ton else to say. It was acceptable.

23:52 And again, I think the biggest takeaway that I've just been really thinking about the last month or so is basically any model we've tested in that time span. Were we to have seen it a year

24:02 ago, it would have been state-of-the-art. And this is no different than that. So, the rate of improvement has really been something special. And it's just cool to stop and

24:10 think about that every once in a while and go back and see what models were doing a year or two ago because then things like this become far more impressive. So, with that, that's going

24:19 to conclude our first look and test of Ling 3.1 Flash. I am also testing that Strata thing that folks have been talking about a lot. There's a system behind me if you can see it. And uh it's

24:31 up on the screen right there and I'm using it with the Quen 3.8 next at a 3bit quant and it's actually I mean it was showing likeund and something tokens per second and it's

24:43 offloaded as well. So, it should be kind of interesting. That'll be a video that should come pretty soon. All right, if you have any questions, feel free to leave them in the comments unless

24:52 they're mean. And thanks for watching.