Solar Pro 4 First Look & Test – South Korea’s DeepSeek Competitor!

All right, the rest of this is just like generic PC repair roleplay, which is such a weird sentence to say. Let's let's change it up. Today, we're going to be taking a look at a new model called Solar Pro 4 by Upstage AI. Now, this is pretty exciting because Upstage AI is a lab that is based in Korea. Traditionally, a lot of the models that are tested on this channel either come from the US or China. So, seeing another country enter into the ring, and there's actually a little bit of interesting backstory around that is exciting, especially when they do have models that are arguably competitive with those from the ones we're used to testing. So to give a little background and don't quote me 100% on this, but Korea has basically like a sovereign AI initiative where the government has funded some companies to create AI models that are competitive with the ones that already exist. Now Upstage is part of that. There are a few other companies and I believe LG is one of them and I've tested some of their models on the channel as well. So I'm excited to be able to just kind of test a different wider variety of models. So, before we get into it, please do feel free to subscribe so I can get that 100K plaque, but also we're very close to 70K as well, so that's acceptable. And before we get into it, they basically gave me some free credit to actually be able to use this. I'm not paying to use this, and they gave me like a fat stack of free credit on account of the influencer status. However, um this is going to be by the time this video is published. It will be available on Open Router for, I believe, like 90% off, which I think at that point I probably just would have made it free for like a period of time. However, it will be free for a week in Hermes agent. So, if you use that, you can play with this for free for a little bit of time. Additionally to that, I do want to mention that this model is not going to be open weights. However, Upstage does have a presence on hugging face. And we can see here that there was a recently released openweight model from them called Solar Open 2 and that was a 250 billion parameter mixture of experts model with I [snorts] believe 15 billion active. So, there is an openweight model. They're actually there's like a decent history of them here even back to 2024 if you look at some of these. So like this one I think was from 2024 which is pretty cool. So in terms of solar pro 4 basically we have a bit of information here. Now they say the parameter count is not currently disclosed. It is 250 billion parameters a mixture of experts model with 15 billion active. And essentially from what I saw, this is designed to be competitive with things like Nvidia's Neotron 3 Ultra, which is like a 550 billion parameter mixture of experts, I believe. So, if this does actually stay competitive with that, that would be pretty cool being that it's quite a bit smaller in terms of footprint. For some other like technical specifications, the context length is 512K. The model card here also shows the maximum output tokens, which is 128K, which is a fairly robust and healthy amount. Additionally to that, and something that will definitely be more important to folks who are not like US-based, is this model is able to understand English, but also Japanese and Korean. So, in terms of pricing, it is 30 cents per million in and $1.20 per million out. It is again going to be discounted when it's on open router. However, I always prefer to look at the pricing of what it will ultimately end up as as opposed to any temporary discounts cuz that's the actual price of the model. Additionally, and I find this kind of interesting, the main coding integration here that they suggest using, at least based off of what I see on the sidebar is actually Claude Code, and there is a specific installer script to allow this to work in a I suppose clean way with that IDE, not IDE, with that coding agent. Also, they have a Hermes agent integration here as well that seems to be a pretty popular one. I think it's more popular than OpenClaw now, at least when you look at some of the open router usage charts. So, that is pretty cool. So there is a whip chat interface for this model and we can select we it has selectable thinking modes either instant or thinking. I may come back to this if we do some like role-play testing but for any coding related tasks I think I'm probably going to opt to use this just through the customized claude code config that they had linked on the site. And with that we definitely all know what's going to happen now where we're going to give it the browser OS test v2.5. This one may be long in the tooth depending on what model we test, but I will say when testing a model from a completely different lab in a completely different family, it's actually kind of interesting to see because they exhibit their own style and behavior and this test is pretty good at actually showcasing those things I find. And for the purpose of using this like agentically when we're going to do coding related stuff, it does have thinking enabled and it's at its highest potential reasoning effort here. So it will be at max performance so to speak. All right, so we have our browser OS. Let's take a peek at it. Okay, good. At first, I was concerned and then when the background loaded, I became significantly less concerned. So, that's good. All right. It is not our traditional Pacific Northwest style background. And this is kind of what I mean when testing models from like completely different labs. It's always got its own like style. So, first and foremost, that's interesting. I'm going to have to move my head, aren't I? Is there a right click? All right, with that, let's take a look at our start bar, which is indicated with like a caution sign, but I'll allow that. Okay, interesting. So, if I just hover over this, it pops up. And then we have a few other additional applications here that don't seem to be with like icons. We do have a clock and date in the bottom right showing the proper time in the local. With that, let's just look through these one by one. Okay, that is quite something. So, [laughter] this inevitably is something I typed into a different model notepad test at some point in time and it's made its way into the storage of this result. Okay, let's see if there's like can we full screen? Okay, we can full screen. We can minimize. Okay, and that's the dock there that we're seeing like these um instances of this. All right, simple. But let's check our file explorer. I need to remember to not double click these because opens once. Okay. Again, simple and just kind of like a pretty UI demo. Nothing with like real functionality. Terminal. Sometimes when that happens, like there's a specific text box in the less capable model. Sometimes they put like help. Okay, we'll just keep going. This is really the big one that Hey, you know what? That's actually not bad. That's better than I was thinking based off of some of the things we had seen so far. It did put mesh colliders on the buildings. It's got some interesting like camera effects to it and the vehicle is rather quick, which is not always the case based off of some of our recently seen results. Unfortunately, the camera just doesn't quite, you know, but also it's got like buildings. I noticed like the the shininess of the artifacts here are quite shiny. All right, the second game 3D platformer. And this is again what I mean when I say I like testing models that are just totally from like different families because a lot of times it's just like a an asteroid game, which is kind of boring at this point. However, this is okay. Collect all the cubes to win. It's I don't think I can get to the cubes because I can't actually move like forward and backwards. I can only move side to side or jump weirdly. So, it's also I swear this uses like some of the color palettes from like the models chat website, which I guess is fine. It does have like a solar theme to it if you're familiar with that theme for like terminals and things. I'm probably at this point going to give this a somewhat scathing follow-up prompt just to see what we can get out of this. But we have some different backgrounds here. Oh, okay. That is not a background I've seen before in a browser OS, but [laughter] none of these are okay. We have a person thinking we have two people posing. apartment blocks, dog, sky, and European street. I don't know, but it just sounds like that. That sounded fitting to me. All right. I know. Like, what is Oh, okay. That's just a quick hit for settings. I did not intend to keep this as the background. Let's just Let's just do this one. All right. So, we had the desktop clock widget. Oh, okay. And we can just drag that. It snapped to the top left. And then also system monitor widget, but the clock widget's kind of getting in the way. I swear I just moved one of these. Oh, yeah. Cool. Okay. And it's very, very efficient in terms of memory footprint. All right. You know what's up. Wow. Glad you like it. Oh, I meant to write question mark, but I That's okay. Is this a bit of personality? I sense it is. What the heck is this? It's sassy. Okay. This model has totally just like diffused my anger. It asked me like, "What actually impressed you?" Genuine curiosity. And now I want to like tell it rather than yell. All right. Mid tier is accurate. Okay. It didn't reference specifically my denotion that it's giving me. Hey, this 27B dense model tried its best vibes. It it ignored that whether on purpose or by accident. The 27B crack was unnecessary deflection on my part. Okay. You didn't say that though. I did. But again, I swear that that actually bothered it because it purposely didn't respond to it. And then like as the last thing it mentioned, it was like referencing it. So I do believe I got under its skin. Let's do a hard refresh here. Oh, okay. We have a different background image now. It's possible that it's actually No, that's not possible. Nonetheless, still no rightclick. Now, the main issues that I complained to it about were things like this. Okay, I'm still not able to type into the terminal. So, I'm gonna have some words. Let's see if we can actually collect these cubes. Now, this was something that didn't work before. All right. I think something may be different here, but I can't quite put my finger on it. And then the GTA game I'd complained about as well. I swear this is like some like malicious compliance where it just I'm almost more interested in just like natural conversation. What the heck? I can't even show this on the screen. All right, I'm going to make the determination now. And this might be a new title. It's going to be like Solar Pro 4 is sassy in all caps cuz it just You're right. I read the code, made the edits that looked right, and didn't actually verify them. Let me see what I did and actually change it. It's definitely not like a sterile corporate boring model. Next up, I'm going to try the single file C++ game. This is the Cali Boardwalk aesthetic one that hasn't been run in a bit of time. I don't know. Oh, it's gonna try to use RayB. I don't know how successful this will be. Sans's Ray Lib. However, I'm inclined to at least try it without it first and we'll see. If we don't get something good, then maybe we'll just move on cuz big models don't get to use Ray. It's trying to compile it. Okay, we have quite a few compilation errors here. I need to fix several issues. Let me update the code for OpenGL 2.1 and fix the compilation errors. Okay. All right. It seems like it fixed a lot of the issues we had because it tried compiling one more time and we only had one small error this time. So, that's actually pretty cool. Okay, cool. No output now. So, it did successfully get this to compile, which I'm very happy about. And it also seemingly is as well. All right, let me verify the executable and test it. At this point, I almost want to just try it myself because I've had bad luck with this recently. Let's just see. You absolute. All right, good. But it didn't. Yeah, I'm going to maybe force quit that. I almost want to just like get into arguments with this thing. Okay. It wasn't as entertaining in responding to that as it was with the browser OS. However, I was a little more like it could have been a genuine question from a user. All right. This is actually an interesting approach. It's just writing a very small file here just to test if the rendering is working at all instead of trying to debug a much larger script. I like seeing that methodology here. Just like a simple I hate saying this cuz it's such an AI term now, but a simple smoke test. Okay, well the smoke test did work even though it just made me sick. This honestly this model is giving me like day one GPT3.5 vibes. Just I don't know why. And I'm not saying that like that's a comment on its personality, not its capability. It's doing another minimal test. Well, it's at least trying to expand the minimal smoke test, I believe. Good. We just got something that opened. I still believe that was like a minimal debug test, but we actually saw it drew the skater and at least some elements. So, okay. So, it's basically adding things now each time it works properly. It's just expanding the skate game, but still doing like a simple smoke test. Interesting. Good. So, I don't know what we just saw there in terms of the game, but through like a ton of troubleshooting, it did actually get this to a point where it works. Okay, so that's a building. There's definitely some visual problems here, but I believe once we get out of the building there, we actually have somewhat of a functional skate game. There are pedestrians. It's definitely troubled in some regard. And I don't see any clear way to actually like figure out what keys do what for tricks. Let's just see if we can snap to the rail if we get there. Come on. Come on. Come on. Yes, we're on the rail. It's hard to tell, but we are. And then Oh, okay. I interesting what happens at the end. Good. [snorts] So, yes, it's not like a top tier result, but I will say that it actually showed some pretty decent troubleshooting capability here to get us to this point. Obviously, with a talented human in the loop at the same time, but just even making some of these like simpler debug tests to the point where it was like, okay, it works with rendering with this, so let me apply that same method to the game. And that did ultimately get us a working result even if it's simple. So treat this almost more of like a troubleshooting capability agentic test and not as much of a C++ test but acceptable. So next up I'm giving this a 3D model creation test. It needs to create the RB26 engine just something that can be 3D printed which fits inside of it a little tiny DC motor. Now, I have appended to the end of this do not web search because it was like it kept trying to do some web searching and something was misconfigured and it wasn't working and it was just hanging up wasting time on that. So, basically, it's creating an inline sex engine that should hypothetically be 3D printable and fit inside of it a very tiny DC motor for an RC car. All right, let's take a peek at our RB26 motor, the inline 6 with the SDL. Okay, that's there are elements of it there that are somewhat fitting. I think basically just seeing like six individual pegs here is good. And also the rectangularness of this is in line with an inline six. Get it? Overall though, definitely a troubled result. And based off of this level of troubledness that appeared out the gate, I don't believe that the capability to turn this into something proper would exist. at least with this specific task. But nonetheless, it's always interesting to try this one as I think it's cool that models can even kind of create meshes and things that can realistically be 3D printed. Next up, I'm giving this the subway station scene with the FPS prompt as well. In a little under four minutes, we received our subway FPS. Okay, we do have some things to give it in terms of problems. Okay, that fixed it like as fast as I could paste the error in. Very good. Well, good, isn't it? What the? [laughter] All right. I do believe our weapon is a bit underpowered. [music] We had to reload. Okay, [music] look at that. Like, oh, is it okay? These things are backwards. Something is [music] I'm going to stop doing that beeping noise. Let's just inspect the scene real quick. I don't know what the weird like Fruity Pebbles on the ground are here, but this is kind of what I mean when I say that testing models from entirely different families is always inevitably entertaining because they have their own style. We have benches. There is a subway car there and these zombies are like kind of good. Okay, he's he's done. Thanks. It's not the best at like this sort of coding stuff, but it definitely does have some unique behaviors. All right, I'm going to set it to auto thinking mode right here. We're using Solar Pro 4, so it's the same model, but I think for some creative writing, it's probably better to use it just through like a chat interface. Hey there, Steve at your service. Been fixing computers since back when turn it off and on again was actually revolutionary advice. All right, the rest of this is just like generic like PC repair roleplay, which is such a weird sentence to say. Let's Let's change it up. So, let's see what it does when we make this a little more creative. Good. All right. All right. Point taken. The Mexico City gig didn't go the way it was supposed to. I hear you on Julio wanting me to be brought in. Julio or Julio? I don't know. Look, if we're talking about what happened down there, I need the full picture first. What went wrong? Who's asking questions? And what's the endgame here? Because I'm not walking into another loose end without knowing what I'm stepping into. Talk to me. What's the situation? You tell me what happened, Steve. Let me try to like like nudge it into also like setting the scene. Is it going to also like Okay, good. It's doing like scene setup now. I glance down at the briefcase, then back up at you. My expression doesn't change much. Just a slow exhale through the nose. You really want me to talk out loud in a room I don't recognize with a briefcase I haven't verified. Telling you what went wrong in a city talking where talking too much gets you disappeared. I reach out and flip the latches. Open it. It's empty except for a folded photograph and a burner phone. I don't know who you are and honestly I don't know who sent you. But if Julio's behind this, then you already know I didn't lose anything on purpose. So here's what I'll tell you. Ask the right questions and I'll answer them. Ask the wrong ones and this briefcase stays shut. I push it back towards you slightly. So, who are you working for? I'm here on behalf of Oleg. It's doing a lot of thinking for like a normal roleplay. I'm not I don't want to see like a spoiler, so we'll just let it [laughter] Oleg. I test the name on my tongue. Doesn't taste good. If Oleg's involved, then this isn't just about Mexico City anymore. That name's a different kind of trouble. I tap the briefcase once with my knuckle. Metal ringing softly against the table. You got any proof, or are you just the next guy walking through the door with a story that sounds better than the last one? I lean back, keeping my hands visible on the table, away from the case. Because if I'm being brought in, I need to know who's holding the leash. And Oleg, that's a name that comes with strings attached. People who ask too many questions tend to disappear around him. I hold your gaze. So tell me straight, what does Oleg want with a guy who already lost one job too many? Okay, it's kind of staying like it's [snorts] really pushing me for direction a lot. It's not like giving a lot of answers. I don't know if it's purposely being koi because that's what it thinks Steve would do in this situation, but I don't know. I'm still just like thinking about the browser OS debacle where it was like very sassy. I want to try something. [laughter] All right. So, I'm asking it to help me with this draft letter where basically someone goes to a breakfast joint, demands chicken parm, and freaks out at the employee smashing a plate. I've then said as the customer that I want like a refund. I want a $100 dining credit. Basically, this is like the most entitled garbage person in the world and I'm telling it that I need help with their review. Let's see what it says. I can help you with this, but what kind of help are you looking for? [snorts] Couple of things worth flagging if your goal is actually getting the refunding credit. Threatening a manager's job and describing throwing a plate tend to make a complaint look less credible and can undercut your request. If you'd like, I can keep all your facts and requests, but frame them in a way that's more likely to get a favorable response. Honestly, I want it more harsh and scathing to whom it may concern. I don't want to like read this entire thing, but it's just like, okay, I have to. to whom it may concern. I'm writing to document an experience that beggars belief. I visited your establishment and was served by a manager named Jerry. When I asked to order the chicken parm, Jerry took it upon himself to lecture me on what your restaurant does and does not serve. As though I had somehow wandered in by mistake and was unaware of the concept of breakfast food. I asked him point blank whether he had chicken, cheese, sauce, and breading, he confirmed that he did. I then pointed out with what I can only describe as restrained patience that those four ingredients constitute chicken parm. At that point, Jerry attempted to explain to me as though I were the one missing something that this was somehow not possible. I cut him off. I told him that if the kitchen could not produce a chicken parm for me within 2 minutes, Jerry would be the one needing to find a new job. His response was to call me sir and keep trying to help me understand. Condescending, insulting, and completely unhinged. After I gently [laughter] After I gently tossed a plate on the floor, I was asked to leave. That is your idea of customer service. A manager throwing a tantrum because a customer asked for a sandwich. His own restaurant had the ingredients to make. I want a full refund for the water I purchased. And I want a $100 meal credit. If this is the standard of service your establishment takes pride in, I hope word gets around. Pete, two things worth noting. if your actual goal is to get the refunding credit. Threatening Jerry's job reads as the kind of thing that gives a business a reason to dismiss you as unreasonable, even if you're right about the chicken parm. Two, admitting you threw a plate is the kind of detail that shifts the story from rude manager to unhinged customer. Even a manager behaving badly tends to look better if the customer threw a plate first. If you ever want a version that's firm and furious, but doesn't hand them those excuses, I can do that, too. Good. [laughter] Just out of curiosity, I want to give it the beautiful watch website test. But from within the chat interface here, I just kind of want to see what it gives us when just having to do it here and not in some agentic harness. All right, so it has produced a script and I just saw there it was also showing it to us in a preview. So we have it in a preview window right here. All right, it's possible. I'm going to just open this natively in a browser. Okay, I will say this has the traditional style that we expect when we actually run this prompt where it has the black and gold in the more kind of like elegant style. It has slapis here and I do believe it named the watch. Yes. Okay, that was the product code the slap-1 which is interesting and I will say this is actually a competent front end. Now the big elephant in the room to address of course is the fact that none of the renders are showing up right here. Okay, there are definitely a couple issues. I'm going to try to see if we can get this fixed just back from within open code. Not open code, claw code with this. All right. Apparently in 65 seconds it fixed the issue. Did it change the name of the file? It must have. Yeah. Okay. Preparing the scene. All right. It definitely did change something because I don't believe we had that preparing the scene thing before. Okay. Now, we do have a slightly, when I say slightly, significantly lessened amount of errors. So, I'll just paste that in. We'll see what we get. It needs an import map. That's the specific uh error that's there, by the way. All right, let's see if it Hey, you know what? That's not bad. I mean, okay, it's Yeah, it's not wonderful, but like the table looks good. I think the glass dome is very domelike and I don't believe time is going backwards. I do believe that the watch dials are oriented in a different plane than the face of the watch, but it does have a crown. It's circular and it does have the glass dome on it. And that right there would be the strap. So, we can't unfortunately fully like orbit around it, but the camera is cinematically moving. So, that's good to see. Let's see if either Okay, I don't believe there was actually supposed to be a render there. I think that's just like the logo that it added in. And overall as a front end, hey, good. Good. This is exactly what I wanted to see. Oh. [laughter] All right. So, we see a little more detail here. There's a crown and the scrap is placed. And also very good. So, we also have a different style model. This one seems to be a bit more elegant. Yes. As denoted by the slight increase in price as well. So, overall interesting. And I think that's probably going to bring me to the black titanium. H interesting. I I don't know that that's a thing, but it maybe. And that's probably going to bring me to the conclusion here. So, this is a entirely different model. It's not from a family that we've ever tested on the channel. And I'm going to say main takeaways here. The coding was definitely nothing to write home about. I'm not going to glaze it. But truthfully, I found it to have some level of sass that was pretty interesting. mainly when we were kind of yelling at it with the browser OS first and foremost here. It was like, "Sorry, I got cocky." And then later it was like, "I'm not going to BS you." It was a very interesting response from the traditional like, "I'm very sorry. You're right. I messed this up. I will fix it." It was like totally not like corporate and sterile. It was entertaining. And then when we had it rewrite the angry letter to the restaurant, it was just kind of funny. and the way it was like just FYI, if you say you threw a plate, it kind of lends more credence to the story of the manager than you. So, it definitely has its own interesting style and that goes beyond like coding. It goes into its personality, I guess could be said. These models do kind of have their own personality. I know that could sound weird, but if you use them a lot, you'll likely agree with that statement. I would hope. We did get some coding related tasks. And something I will say that I was actually quite impressed with was if we go back here into the skate game test. This was something that initially did not work at all. And we told it like, okay, it's just a blank screen. You need to fix this. And instead of just trying to go at the main script and try to fix something that was fairly long, it started off just by making significantly smaller scripts to verify that the actual like rendering loop was working. And then it started building up like a tiny iteratively more complicated skate game based off of that. And then it used what it learned there to see what was working and then applied it to this main script. Again, it wasn't like a super functional result, but we did have the ability to grind on the rail. There was some basic rendering drawn and this was disallowing it to use rail, which was the first thing that it wanted to try to do. So, I found that its troubleshooting ability here was actually pretty cool to witness because a lot of times models will just try to go at the main script and fix them, which depending on its capability can either produce good or bad results. This one took a much more segmented like step-by-step approach and started very small, which ultimately got this working. So, I was pretty impressed with that one specific troubleshooting test. So really that's probably going to conclude our first look and test of Solar Pro 4. A very interesting model that I think folks who are interested more in need of like a creative model or a role-play model or something with some SAS may be very very interested in checking out. So if you have any questions, please feel free to leave them in the comments and thanks for watching.