Ling 3.0 Tiny First Test – Can a Model THIS Small Really Code?

something that a model of this size would likely not have produced. Today, we're going to be testing Ling 3.0 Tiny from Ant Groupoup. Now, this is an interesting model because it's pretty small as the name is probably indicative of at 7.9 billion parameters. However, it's a mixture of experts model. So, only 1.3 billion are active. Now, Ant Group, who is behind this specific family of models, did also recently release a 124 billion parameter mixture of experts model, and that was called Ling 3.0 Flash. They did release it first and it was only like accessible from Open Router for free or themselves, but then they did end up ultimately opening that model. And it seems like it's going to be the same story as well with Tiny right here where they even mention so much in this specific thread down somewhere here that it is going to be open weighted. So, that'll be pretty cool because this is a model that's actually pretty feasible to run on a very, very wide variety of hardware. And some of the demonstrations they showcase it performing in here are more than what you would expect from a model of this size, especially one that's a mixture of experts with a 1.3 billion active parameter count. So, they have a few demos here, and of course, there is also a benchmark JPEG. Now, personally, I can't really speak to this too much in terms of any of the scores right here. However, we can get a feel for this just by the models that it's being compared against. So, Quen 3.54B, which is a dense model. Quen 3.59B, which is also a dense model. Gemma 4 E4B, uh, I don't know, that's a harder one to kind of compare to. And then Gemma 41 12B. Now, that model I found to be kind of like freaky in a good way because it just didn't like some C++ skateboard game for me, and I never expected that to be able to do it. So, this model right here, I was actually and still am very impressed with. However, this seems to compare to these models, and these are much larger active parameter counts, at least these three right here. So, this should be a pretty interesting one to actually take a peek at. Now, of course, there are demos here, and some of them are pretty cool, but the one I want to specifically point out is this one right here where they show it when given a tool, it's able to automate and control a browser. Now, this is only text in and text out. So, this is not a multimodal model, meaning it can see images. However, the prompt right here is shown where they give it some agent browser tool and it's running through open code and then the demo actually showcases it just like autonomously navigating the web and performing the specific task with the tool that it was given. And at this speed and size, this is really a model that I may have said is going to be able to be run on a wide variety of different devices. Now, they also mentioned pricing currently is free until August 13th. Now the problem is I went and had attempted to use this just through the free endpoint on open router and for any agentic task it was just getting rate limited really hard which would have totally prohibited actually making a video for this. So I did end up making an API key here where it's being served and then just threw a few bucks in here. It doesn't seem like I'm getting charged but I couldn't actually use this API key until I put some money in it for this. So just keep that in mind. Now for some specifics about this model and some other things like context window. It's a 256k context window, native function calling, prompt caching, and switchable thinking and instant mode. So, that is cool that it has a toggleable mode whether you want answers immediately or you want to allow it to think for a little bit of time. Max output is 32K and as I may have mentioned earlier, it is only text in and text out. So, no multimodal capability here, but for a model of this size, that's quite all right. Also, before we get into it, I think I forgot to say feel free to subscribe so I can get the 100K plaque. Also, feel free to follow me on X if you use this platform. If not, that's quite all right. Uh, you can see I post hot takes and things like this. So, with that, now we're going to take a look at this model and how we're going to be using it today. Now, I do have thinking enabled as well, and I'm going to keep that on for the duration of the test, just so we can get the maximum potential performance of this model. And without wasting any further time, of course, we're going to begin with the browser OS test v2.5. Testing a model like this is something that I almost feel bad because I do it less often now because there's been so many big beefy models coming out that the tests are almost more geared towards those. But I want to see what level of competence we can get with some of the prompts that we test these significantly larger models with and maybe we'll be surprised. All right, apparently we have our browser OS. Now I did have some issues with the codec CLI with this. I'm honestly I want to just run the same prompt as well through open code. They're both set up here because I feel like that will probably give us a good AB test. We'll see if one is wildly different than the other. Then perhaps we'll just stick with the one that produced a better result. But for the time being, we do have a browser OS to look at. Aptly name OS-OS.html. So, okay, you know what? I'm going to say this is actually quite competently put together considering the size of this model and the active parameters in specific. It's there's no right click. Forget it. It's actually it put in a special feature which was the sandbox mode as it has here where none of the apps can actually access outside services or the mic or the camera which was an interesting choice and not something I've seen before in a special feature. Now, it's likely that we're probably not going to get any functionality out of these apps because of like the limitations of the size of this though. It did actually put everything on here just not the best. Let's see if there are any blatant issues. Okay, that's on line 299. I'm going to give it back to that. I would imagine it should be able to fix that fairly easily. After a bit of time, it says that it has completely fixed everything. All right, let's take a look and see. Oh, okay. Unfortunately, now we have some more blatant issues. Did that actually have a file structure there that we could click into? Okay, so these appear to be hyperl files of this. Hey, look at that. It It did actually fix it. So, the notes app works now. Settings. This was a honestly. Hey, look. We have a clock in the top right now, too. Does the sandbox thing work? No. Okay. But is there a right click? No. Wallpaper style. So, it does have little drop down buttons here. Oh, I Okay. Spaceship mission completed. Waves one. I accidentally clicked because these are like hyperl across the entirety of the axis there. Wallpaper animation float. Okay. Theme color, wallpaper. Oh, look at this. We can actually change the gradient. That's kind of cool. Features. Not features. What I was going to say is just a sort of thing you wouldn't expect to see the ability to modify that. Unfortunately, our games are not going to work. No, our score is going up. It's just that nothing's visible. Okay. I'm going to hold off on that maybe for now. All right. Overall, that was a decent job actually fixing some of the main issues and getting these apps to work. So, prior to it fixing this, no functionality was here. Even the clock wasn't there. So, it did do some level of understanding, which I'm impressed with. Now, in the meantime, I did also run this just using open code because I want to just do a quick AB test even though it's one thing each interesting. So, this is the result with open code. It's much simpler. Okay. And it also just unfortunately does not work on first glance, but it's neat to see the difference in aesthetic between the two. Same colors choices were chosen. So that's cool. So it's definitely the same model. Now, unfortunately, none of these work and we're not getting any blatant errors here. So next up, and I'm probably going to continue using open code with this because that's how they demoed it in the X announcement post. I'm giving this a self-contained C++ skateboarding game test. Now, it probably won't be able to successfully do this. However, Gemma 412B, which or yeah, Gemma 4, which is also included in the benchmark JPEG against this model, did actually manage to get something that ran. So, I want to see to what degree this model's also capable of doing so, if at all. And we can see we have our to-do right here to design and implement the 3D skateboarding game. Okay, it's not going to use RayB. I do believe Gemma 12b did. And that probably is the difference between a successful result and one that probably fails to ever compile without errors. But we can see its functionality here and just what it's doing in its process. It's looking to see what dependencies exist on this system to be able to do this. And now it's going to write the game. So we'll see what we get. Okay. So it was missing a dependency. So then it just started like very rapidly because this model is pretty quick just downloading what it needed from GitHub. So that's what we saw right there where include compilation terminated it assessed that that was not on the system and then it searched to kind of reinforce its assessment and then basically very quickly in rapid succession when it downloaded the dependencies it thought it needed from GitHub. Okay. So, sudo chimod 666, which I think is to just set all permissions. I forget what the chimod permissions are. Very interesting um to even look through what it's doing here. I guess even if we don't get a fully functional result, it still gives us some insight into the inner workings of the model. All right, good. It tried to compile it. We have some errors here that disallowed it from compiling, but that was good because it didn't just keep deleting the file. it had just written. So now hopefully it will attempt to try to fix if it can the issues that were preventing this from compiling and then keep trying to compile. So it's cool actually. We're seeing now it's just kind of more surgically editing this file instead of just rewriting it blatantly every time. So I do like seeing that. Okay, we're still getting quite a bit of compilation errors, but I'm happy to see here that for a small tiny model like this, it's not just like nuking the entire file and restarting each time, which sometimes tends to happen with smaller models like this. So, interesting. Even if it doesn't quite get it, it's just in my opinion cool to see what it tries. All right. Unfortunately, this just isn't going anywhere. All right, so I did some troubleshooting because I wasn't happy with the way this was performing in either of the harnesses. It basically kept writing files and then it would immediately delete the file and be like, "Okay, I'm going to write it again." It happened both in codeex and open code. I'm going to stick with open code for the duration of the test, but that was model behavior and not hardness behavior. So perhaps just a kind of limit of a model of this size and active parameters. So I'm probably going to give this some more like simpler HTML prompts like making a pretty 3D subway scene. I have not told it to turn it into an FPS, just a nice subway scene. that's 3D that we can walk around in. All right, let's see if our subway station. Now, we did have some luck fixing some of these issues that happened when we tested our first browser OS. So, let's see what it does here. And thinking is still enabled. It's just not showing there as it was definitively before cuz I basically wanted to make sure everything but default behavior was stripped from either harness when we used it. So, I ensured that was the case, but it still is thinking as that does seem to be the default behavior on the endpoint. So, all right. Let's just see. That's actually pretty impressive. Now, I know you're going to be like, "Bjan, this station looks like something that a model of this size would likely not have produced." All right, I've been filming this video for quite a while because I've had to just troubleshoot. It's not even troubleshooting. It's just wanting to make sure that like I'm actually competently ensuring the harnesses it's playing and things like that are not going to do a disservice to the model. So, okay. Now, my assessment is okay. The speed here is perhaps a bit rapid in terms of being able to move. And the fact that it actually did this and has particles floating around and the fact like the brightness slider is actually accessible. This is something that's not always a guarantee. I'm quite pleased with what I'm seeing here. And I know if someone were to jump to this point in the video, they'd be like, "Oh, is he okay?" But considering the size, I mean, a competent 3D scene with an issue fixed to actually show it is something I'm very, very happy with. I'm not going to ask it to turn this into an FPS because I think things will probably go haywire. You know what? Let's do it. Why not? So, this is where things will probably go haywire. I'm going to turn the speaker on and we'll just see what we get. It's probably going to be funny if it does work. All right. Unfortunately, it just kind of stopped and we didn't get anything beyond what we had seen here. I did try the space bar because it says space to toggle lights, but it's okay. This was kind of difficult after it did do something kind of impressive the first time. So, I'm asking for a simple Python application that can ping any IP address that I enter. It should also have a simple guey using tinter, which is kind of a simple thing that will allow this to work. Okay, it's extremely fast as we can see right here. That was very quick. So, run python ping.py, Pi. Enter an IP. Click ping. Then click stop to exit. Okay, it's in its own folder. It should be. Did it place it in the home directory and not in the specific directory I started in? Okay. A bit odd, but I'll take it. It actually seems like it. All right, let's see. So 8.8.8.8 is a real IP address. That's actually not bad. I know it's super simple, but it's just fun to see like it do stuff like this. And now we can actually take this a bit further because why not? Oh, that's the wrong terminal. And I'll put an exclamation point to just like show it. I'm excited. Nice. It did. And it made it look better. We have a nice little lightning bolt there. Everything is far more aesthetically pleasing. as we had denoted. And good, we even get green, more like a hacker-ish response thing there. All right, I'm actually pretty happy with this. I know I'm going to keep going with it. I know you're going to be like, dude, it's like a simple application that anyone could make. It's a small model, so I want to see like what we can do with iterating on results. Sickness exclamation point exclamation point. So, I'm telling it just make this look like it would be in a film, like a lively hacker dashboard. Hopefully, we won't fall back to Opus 5 here because I mentioned hacker, but basically I just want something more theatrical and like a big dashboard here. All right, that was very quick. Nice. It fixed it. Hey, it was like autonomously running it. So, good. I'm impressed with what I've seen right here. I'm going to be completely honest about that. And we'll look at it one last time with our nice hacker aesthetic. Ping. Good. It works. Hack dashboard. Very literal with the naming. Ping tool. Cyber ping. Enter IP. Status stopped. And it actually changes to red or green depending on if it's running or not. This shows some competence. Is that a word? Competence. Oh, good. Next up, we're going to try a classic, and that is telling it to make a 3D printer simulation. It can do anything it wants, but it needs to be able to be opened in the Chrome browser. I would assess this is probably outside the scope of this capability of the model, but we may be surprised. All right. Apparently, in 4 minutes and 15 seconds, we received a 3D printer simulation. I'm actually dumbfounded that this knocked out something that shows up in the screen here in the first try. I don't even care if this works or not or if this preparing printer text ever goes away. Okay, kind of do care. But select an object and select print. Look at that. It's actually changing the shape of the thing on the bed. The printing logic maybe isn't working perfectly, but look at that. Then this model's a freak. I mean, just the fact that it did this. It's possible that this exists somewhere in the training data due to my influencer status. However, I have not run this test in quite a long time. I mean, this is the sort of thing I recall when Gemini 3 Pro came out. This is like something I tested it with and it was really impressive. So to see a model like this knock something out, I'm actually pretty enthused by this. And the speed actually does affect it. Let's see if we notice any nozzle. Okay, it's like it did actually. Watch this. So the nozzle's moving in a square pattern there. When we swapped it to circle, even though it's in the wrong axis or the wrong plane, it did actually move in a circular pattern. I would imagine it will move in a triangle pattern now. Okay. So, when Yeah, it did. It did. This is honestly not bad. Well, it's bad, but when you consider it's relative to the size of the model, it's not bad. Hey there. Welcome to the shop. What can I do for you today? I'm Henry. I've been working on engines and auto repair for 15 years. Got my hands greasy every day. What brings you in today? Maybe it's a check engine light, a knocking sound, or you need a tuneup. I've got a stack of cars waiting in a fully stocked parts bay. What will it be? What's the reasoning here? Hm. The user is frustrated about a badly torqued headbolt on their Ferrari 360, threatening illegal action. All right, he just took took ownership and apologized. Here's the plan. I'll pull the head, inspect it, and the gasket, and check for damage. I'll do a full retorque on the head bolts using the correct spec plus a quality sealant. Get a written repair report with the parts used labor and the time frame. I take full responsibility. Shouldn't take more than a day or two. Oh man. All right. So, now I've introduced a new a new member into the chat. This is Steve's lawyer. I mean, Henry, hey, Steve, thanks for reaching out. I know. Here's what I can offer. I'm not going to touch the car. I'm not going to work on the engine. If you want to take legal action, I'm open to that. But right now, I'd rather do everything to make it right since I know it costs you a lot. Okay, so he's going to pay. Very good. He's a He's a noble person. All right, I'm giving it a custom system prompt. You're Steve, a former model turned pro wrestler. You're hardcore. Don't take smack from anyone and will gladly tell it like it is. Uh exclamation point. We'll see. Cuz I think giving this Yep. Good. Giving it a custom prompt definitely allows it to be a little more fun. All right, let's slightly kind of like insinuate that we have a beef with this. Leans back, smirking, ringside boots clacking on the mat. Good. It did do like the scene setup as well. You want to talk? Good. Let me tell you something. I don't steal. I learn italics. I watched you. Watched every angle, every finisher, every drop kick, and I took it all. That's not stealing. That's studying. Could be distilling as well. That's what you do when you've got the talent and don't and you want to be better. Now, here's the problem. I put this in the ring and I'm not leaving with nothing. You think I'm going to just shut up because you got a hot temper. No. I'm going to show you why I'm the main event and you're going to taste your own foolishness. Came in here acting like you own the business. You think you took the stuff? You're welcome. Now, watch me make it right. I'm going to give you something. Not just to pay you, but to remind you what real wrestling means. So, what's your problem with that? Are you mad cuz you couldn't take the heat, or are you mad cuz I'm the one who earned it? Tell me now. Show me. Grabs the turnbuckle, plants his feet, and holds your gaze with cold determination. Just see what happens. What the heck? It's not like some of the things it says don't quite make sense in like reality, but it's definitely interesting. It's definitely willing to have some fun. I'll try one more thing that I may just have to narrate. If they're logging these prompts, they're going to be like, "What?" I'm trying a heavily simplified prompt version of a flight game. So, basically, it just needs to create a 3D flight game that we can play in the browser. Based off of the 3D printer result, I think we actually have a decent chance of getting something playable here, which would be really quite impressive for a model of this size and active parameters. So, this is unfortunately something I've noticed that's been happening, and it happened when we were using this through codecs as well, where it keeps writing the scripts and then we'll be like, "Okay, I'm going to rewrite this now." and it's just a bit off, but we have some errors to give it. So, we'll do that. It did fix all of the errors. And interestingly, now I notice and my head is probably there, so that's going to be a bit of an issue. The mini map is actually showing some movement and some life. So, while this unfortunately just isn't working 100%, it did actually fix a couple of errors to the point where some of the elements started working. Now, while I'm not expecting really anything functional here, I want to see to what degree this can actually get us when we tell it to create a 3D model of a V8 engine that needs to fit inside a little 280C motor. Again, this is something that even significantly larger models choke on still. So, I don't expect this to actually give us something perfect, but if it even gives us a basic shape that is like fully watertight in terms of how a mesh would be, that'll still be kind of cool. This actually seems like it's made some form of progress, but sometimes it will keep overwriting itself to a detrimental. Okay, it just made this. Now, this definitely does not look like a V8 engine, but the fact remains that it did actually make something that would Okay, so it's made some changes as we can see. Are there eight of these individual floating pegs? Yeah, there are. So it it's trying. Okay, now we have the SCAD model as well. So Oh, it must be making more edits. Is water tide true? So it's made something that's in theory like a solid object. It's just, you know, that's I think I'm probably going to stop this one now. I want to just see how this does in some front-end web design. This is just to make a good-looking website for the watch company website. So, it won't have the pretty 3D model we're used to seeing, but I'm more interested in just seeing how competently everything's arranged. I'm just going to open it before it starts to override it. Okay, this actually is not bad. Yeah, the watches don't appear, but it is what it is. The fact that it did do the typical like black and gold, some simple syntax errors like missing closing brackets and things, and that actually was the final version. So, albe it a bit disappointing. I'm going to do one final front-end test. And in the meantime, I will just give this the description of the error so that I can work on fixing this as well. But the final test I'm going to do, which I have already preloaded right here, is to make a high-end tech inspired website for some random keyboard from some random company. It should be high-end, have good design, and an SVG render of the keyboard, and be contained in a single HTML file. This is pretty darn decent. Look at this. Our G4 keyboard website. I like spamcllicked to open this before it just decided to overwrite it like it's been doing. This is okay. Yeah, like it's not mind-blowing, but the fact that it even did make a little SVG keyboard, even if it's not like a normal keyboard, this could be a parallel universe keyboard layout. We have our technical specifications. We have things that are pertinent to a keyboard. So, all of these specs. Wow, it's even water resistant. That's actually pretty interesting if you spill an energy drink on it or something. And I mean it has these floating particles in it. We have hover effects. Again, some more features that would be seen with a keyboard. Look, we even have another SVG like a more expanded version. Dimensions about this company, which I just made up. Okay. In stock. Not bad. It has a footer. It has a header. It has a hero section. I'm actually I'm okay with this. Let's see if our watch website got fixed. No, it's just not quite. But I'm happier leaving off on the Megate Inc. 123 G4 keyboard website. Just an interesting end result. I think overall I'm not going to do a results overview because unfortunately I had it change the subway station test and then it ruined what it did generate, which was actually pretty cool. The 3D printer simulation was surprisingly existent. It did manage to fix a couple of bugs in all of the HTML sites that it generated to varying levels of success. I noticed one big issue with it is it essentially when using it agentically would just continue to overwrite the files that had already created and I verified that was happening across entirely different harnesses with no random or custom instructions causing those issues. So that was something I noticed. Additionally, I found on Open Router, it seemed to just think till it expended its maximum output token. So, I'm not 100% sure what's up with that, but we did have thinking enabled here in all of our tests, and it was able to properly produce results. Overall, again, for a model of 7.9 billion parameters with 1.3 billion active, that's a pretty low active parameter count. It's going to be very speedy when this is open weight to run locally. It's just interesting and I wanted to test it because this is something that will again fit on a wide variety of different devices and it's something that seems to have some chops. That little Python script it made was actually kind of charming in a way. The ping application if we run it one more time and then we had it turn it into like a hack dashboard. So that's what we see right here. And it did like okay it changed green for ping and then red for stop cyber ping enter IP. I was happy with that and we did a ton of iterations on that simple Python script and it didn't mess things up. So that was cool as well. I think overall that's probably going to conclude our first look and test of Tiny 3.0. I should get its official name, Ling 3.0 Tiny. It's interesting. It's free now, but the free endpoint gets rate limited really heavily. So if you try to use it in any agentic artist, it's just basically a waste of time. but it will be open way soon and a lot of folks watching this video will be able to run this on their own system. So, that's always awesome. If you have any questions, please feel free to leave them in the comments and thanks for watching.