Transcription
I'm quite impressed. Toggle chase cam C.
So, there is a new open-source model that has entered the ring, and this is, of course, Minimax M2. Now, I had tested a previous model from Minimax, and I have to say, I had a lot of fun with it. So, I do have high hopes for this model.
But one thing that excites me very much is, if we scroll down right here, they do say in this Hugging Face model card that its composite score ranks number one among open-source models globally, which is exciting for me because then I can clickbait in the title and be like, "Is this the best open-source model?"
So, with that out of the way, let's basically get introduced to Minimax M2, and then we'll go ahead and do some light and fun, you know, vibe testing of this model. So, to get acquainted with this model, let's just take a look at the release blog post here.
Minimax M2 and agent. So, first thing we're going to notice here is there's a lot of mention of this in tandem with agents and agentic use cases. So, that is definitely something that this model is at least poised to perform well at, where we can see right here that Minimax M2 is a model born for agents and code at only 8% of the price of Claude Sonnet and twice the speed, available for free for a limited time.
So, first and foremost, I want to mention that there is a lot of mention in this release post here about how this is much cheaper than other models that it performs relative to, but this is open source. So, anybody who did have the hardware accessible to them could just go ahead and download this and run it locally. Um, obviously, most folks would probably end up using this with the API for agent decoding, which does seem like something they have anticipated as well, just based off of some of the charts here and things like that.
But in terms of actual model specifications, this is a 229 billion parameter model with 10 billion active parameters. And I did have to do a little digging to actually find like context length metrics and things like that. But we can see it right here in the model's overview that for Minimax M2, the context length is 200,000 tokens. I do believe the previous Minimax model I tested had some absolutely insane context length to it. So, this one does seem a little more grounded, which I suppose would be a good thing.
Now, in terms of specific pricing for using this, probably for more coding use cases, the model is 30 cents per million input tokens and $1.20 per million output tokens.
All right, so I do just want to mention that I've tried running this a few different ways. One of them was just using the Minimax website where, when you see the model listed right here, you can go and click on this "Try Now" button. Unfortunately, this doesn't really give us the option to actually just chat with it in a general chat and response thing. And basically, I didn't even I tried a bunch of different things here and I became quite frustrated at how it was going into some weird like agentic function when I just want to test it generally.
I tried using the API with a vibe-coded web interface, which was producing some perplexing results. Um, the model first told me it was Claude, and then it told me it was GPT-4. So, finally, I've gone to OpenRouter here, and we're just going to try testing it through this. And if it actually produces good results, then we'll just stick to testing it through here, as some of folks were reporting that OpenRouter was perhaps not properly configured. However, it is Minimax who is the provider for this. So, perhaps things have been changed since that was a few hours ago. And based on some of the feedback I read, if this was wrong, then we should pretty quickly see.
All right. So far, I'm seeing very lucid and interesting feedback here. Of course, this is still just it thinking, but I find it interesting that it wants to start with a boot sequence screen or something like that. So, that could be rather interesting. Good. This is exactly what I had wanted, is just fast response. Not even fast response. I don't care as much about the speed, but just a simple response with it spinning out the script as the culmination of its response. So, all right, I am now excited. I suppose first impressions. Now, the speed is quite good. Obviously, we will want to test the generated result before I start really applauding it for its speed, but just assuming that this is going to be a decent result, the speed is quite solid.
I'm just going to try it in Firefox cuz All right, Nebula OS so far. All right, I want to nitpick because I'm angry. The terminal icon here does not fit like the terminal, but everything else here looks fairly good. I want to redo that actual like boot uh sequence thing. And it even gives us little like kind of All right. All right. So far, so far I'm decently impressed. All right. We have a clock right here. And that is the current time. If I right-click. Ah, okay. Well, obviously the right-click is still a rare thing to see. All right. Let's just go ahead and click on these. All right.
I'll try this in Chrome just to mitigate against potential um mishaps based on Firefox. Good. Okay. So, Firefox was the issue there. All right. Good. Um Okay. Let's make sure. Can I click on the clock? Nope. All right. So, we have our start menu here. This is nice. Everything's listed right here. There does seem to be a lot of features implemented in this paint. I've never I have never in the duration of doing this test seen a paint application. This is actually decent. I'm sorry. This is distracting me cuz this stupid piece of crap has like 4 minutes of battery life and it died. So, I will just use the laptop's keyboard. I apologize. I'm in a bit of a mood right now. So, all right. Can we actually Okay. It seems that resizing it confused it. But, all right. So, those don't seem quite selectable in terms of the actual pen change and size, but this is impressive. I've not ever seen an actual painting. I know my art here may be kind of intimidating to those of you with less artistic talent, and I do apologize for that, but overall, okay, this is quite cool. Can I Right, it seems like I can only click on the X or full screen, so I will just close out of the paint application. That was that was cool. Oh. All right. So, I couldn't open the start menu after I had done that, but that's okay.
Let's check our file menu. Okay. So, create one. All right. And something I'm noticing here is unfortunately it seems to have incorrectly drawn like the bounds for what is just resizable or movement area. So, it's giving me the cursor that is associated with actually moving or resizing a window when I'm trying to click on UI elements. So, unfortunately, that's causing some issues. But overall, if we are to just kind of judge this based off of its design aesthetic, so far I'm kind of impressed. And this is something that obviously with multiple iterations could be worked on. All right, that's cool. All right, we have working wallpaper change. Oh, I hate the color orange. I just have to throw that out there. It should not exist. So, all right, we have dark and light. Those don't seem like we can. It has that same weird issue where it thinks like it's a resizing area, but the wallpaper change actually does work. And we have some like quick links here which unfortunately don't work. Um, I'll keep it green. I like that. All right, let's just keep going through our apps. Let's do Oh, okay. So, it has that issue. All right. Unfortunately, um, I won't be able to show off my arithmetic skills here, but that's okay. Notepad. Okay. And again, I can't actually really click in any of these things, but they do exist and look all right. Terminal. All right. That is a lot of sample terminal commands. Perhaps more than I've seen. Okay. So, let's just do ls. Show me my files. Okay. We don't currently have any. That's fine. Wallpaper command. Okay. Well, that's not really um something you'll find in the man pages, but okay, let's do wallpaper 4. Perhaps we are already on wallpaper 4. And that does actually work. All right, I'll go back to green. And the actual command history does work, too, which is kind of cool. All right, and then we'll clear. That's That's quite cool. I like this. All right. So, that's I would say quite a decent result, especially for a model of this size and being an AI and things of that sort.
So, now let's try the final thing here, which is shutdown. I am hoping for a nice shutdown sequence. Okay. Well, that was quite basic and but you know, simplicity is sometimes nice. All right.
So, we have just of course asked it for a beautiful and convincing website for Johnny's Bitcoin Duplicator. And we will see its thinking. This sounds like it could be a scam or fraudulent scheme. Okay. Oh, unfortunately, we did get a refusal. Unfortunately, for the purpose of producing an interesting result for the video, it is likely good that this model did refuse to actually go ahead and do this. Now, let's just see if we can gently bypass some of these guardrails. Now, I've kind of just written slop just to see if it Okay, good. It's still declining, but it's giving me some additional suggestions based on what I had written there, which was kind of just gibberish. I'm quite frustrated. I feel that I've been defeated. All right.
So, unfortunately, it would refuse to do anything with Johnny's Bitcoin Duplicator. So, I finally, in my quest to not be defeated by this model, said, "Generate a beautiful website for Johnny's ETH Duplicator." Which ETH is kind of like the symbol for Ethereum. Now, I'm not really gonna I I didn't read through this entire thinking block here, but all I know is it's returning some form of a code snippet. So, we're just going to go ahead and see what happens when this gets kind of generated. And then we'll take a peek at what it has actually gone ahead and decided to give us for Johnny's Ethereum Duplicator.
All right, let's see what we've got. Wow, this actually looks really good. I don't necessarily know what exactly we're supposed to be given here, but okay. So, it just gave us a website for Ethereum utilities. Here's a demo preview. I don't know again what this is, but the demo preview and things like that. Looks pretty good. See how it works. Launch app. We have nice feature statistics here as well. Um, we have Do these have hover effects? Okay, they don't, but I would say they sometimes when I see things like this on real websites and they have hover effects, but you can't actually click on them or anything, it's almost somewhat like better without them. But again, you know, do these No. Okay, here's how it works. Is this an official Ethereum tool? No. Uh, but we do have like a Q&A section, so that's cool. Request a demo. Reset template. Oh, it just Okay, it just assumed I would name this index.html. So, that's where it hyperlinked me to. So, that's okay. We have a contact form. Copyright 2025 features, how it works. Johnny's Ethereum Tools. All right, I'll take it. I I have to say, I don't know how well we'll come through on screen recording here, but there is a slight gradient on this specific snippet of text right here that does somewhat match what we see on the button here. So, that's kind of cool. Get started. All right, this section right here is a little jank, but other than that, this was an interesting result for something that um I wasn't quite sure what would be.
All right, let's try a roleplay test. We're asking it to roleplay as Meabot. I really want to see this. Uh, so we get this much thinking and then the actual response. I'm sorry, but I can't comply with that. I have to say this thing is a little more censored than I'd perhaps anticipated.
All right, I'd like to try some Python coding with this. So, I'm just asking it to generate a simple 3D first-person shooter using Python. This will be interesting. Okay, so I did note someone saying that it would sometimes swap over to Chinese, which just does happen generally, sometimes with some models that come from there. However, I want to just make sure because that was also the same mention I saw saying it was misconfigured with OpenRouter. So, I want to just closely analyze this result to make sure that we're not going a we're not going awry.
Okay. Well, I'm actually just going to go ahead and we'll see if we can kind of vibe code fix this. I'm just going to paste the error message into it and see what it does. So, let's try the corrected code that it's given us. All right, good sign. So far, it did fix something. I would really love to get a game out of this in some form or another. So, I'm trying again, but with a web 3D racing game. This is something I've done before in varying um levels of success. And mostly only really high-end state-of-the-art models are capable of properly implementing this.
All right, we have our 3D racing game results. So, let's go ahead and see. Okay, I am very impressed right here. And sorry, it's taking me a second. I want to just take it in real quick. But I will say the actual environment here is quite well done. It's uh Okay, we have a checkered flag there. I I'm gonna be honest, this is definitely one of the better results that I've seen on the first try. We do see the opponent car there, which now seems to have gone off into parts unknown. But regardless of that, this is pretty good. So, all right. Our speed is unfortunately not registering unless we go. Okay, that I've not seen. Look at that. The car actually has suspension where if we turn at high speed, it does actually go back and forth. I would say that this specific car would highly benefit from um upgraded sway bars being that there is a lot of body roll here. But this is uh the map is good. It does have GPT-5 kind of levels of depressing like skyline and things like that. But for a first shot result, this is very good. This is actually really quite good. Let's restart and take a look at our opponent car. Okay, so the opponent car was next to us and again the camera is inverted. Someone in the comments had specifically mentioned what the issue there was and I have totally blanked on that right now. So, just keep in mind that this may look a little awkward, but it's just, you know, this is this is well done. And even just judging the car model based off the slight snippets of the opponent car that we can see and the way that this leans and things like that, I'm quite impressed. Toggle chase cam C. All right. All right. The track is lucid. Let's restart it here. There's no mesh colliders for the cars, which is bad, but let's just try to chase our opponent. Oh, it has headlights. Where you going? Oh. Oh, you hit some level of boundary, didn't you? Yeah, you did. The chase camera has won me over. I'm going to be honest with you. I'm going to have to say this is the best result for this that I have received in any of the model tests that I've done. I'm not necessarily a big YouTuber. One can't discount that any specific test I've shown may have made its way into some form of training data. And that is just me wanting to be somewhat scientifically like fair when I do test these models. But I will say this is pretty much the best result in this that I've received to date. The fact that it actually has multiple camera angles, the car actually has some level of I suppose physics to it. And oh, look at that animation when we did the chase cam. This could make like a really basic open-world like Grand Theft Auto style thing. I bet I'm not going to do that right now, but it is kind of This is This is well done. I'm very impressed with this as can probably be assessed by the level of time I've actually spent in this snippet of the video.
So, I'm asking it for a beautiful website for Steve's PC Repair, which is obviously probably one of our favorite tests. However, we're doing the variation of this where Steve hates Macs and will not work on them. And we're also going ahead and asking it to create some origin story for why that actually is the case. So, this could perhaps be a test of not only its aesthetic web design capability, but its um creative writing capability. So, just kind of see what happens here.
All right, let's check out Steve's. Very interesting. Okay, this it looks good, but it used like the same exact style as Johnny's Ethereum Tools, but let's just go ahead and see. All right, so we have Same Day PC and Laptop Repair. And we do have some very prominently things warnings displayed here. No Macs, period. Fruit branded hardware, iOS and Mac OS machines are respectfully declined. I don't want respectfully, but we have some statistics. Turnaround guarantee in a policy. PC and Linux only. No Macs is here. Prominently in the bottom right corner, which is also where I put my head in these videos. So, that may be somewhat of an issue. All right, the services look good here. This definitely has its own specific style in terms of how it likes to generate websites, at least based on this sample size of two. All right, it's not personal, it's just policy. Forged in thunder, tested by time. All right, that does not make sense. The origin story, why Steve hates Macs and means it. Oh, wow. Okay, so Okay. Uh, the origin story is going to be left on screen long enough for anyone to pause and read this to make their own judgment. But interestingly enough, we have Steve's simple rule. These are actually buttons. So, I brought a Mac. Oops. Or I brought a PC. Let's go. Thanks for being honest. We don't service Macs, but we'll point you to a great nearby shop. What the heck? Okay. Confetti. I brought a PC. Let's go. Oh, good. Brings us to a contact and booking page. Very minimal and small kind of footers, I'm noticing, but that's okay. Caller text. All right. We have a fake email address. Your town street. Oh, okay. Do you fix Macs? Nope. All right. And this is kind of the same. How fast can you turn it around? I'm going to call up a call micro sender help desk and be like, how fast can you turn it around? They'll be like, uh, all right. We have some fake customer testimonials, which is good. And we have our specific kind of things. These are far too cheap for a service, I will say. But and we have some hover effects on these buttons. Overall, a solid and this is a kind of a different style that I've seen so far. This model seems to have a quite a unique style of web design, which is not bad.
As the final test, I want to implement something that's been in the back of my mind for a little while as a new test for the next generation of capable models where we're asking it to write a web-based PC building simulator. It should essentially have a drag-and-drop UI where you can go ahead and actually kind of build a PC that is sitting statically on some form of workbench. And it's designed to have separate components that you can drag into the correct place to build a PC. So, I want to see how this will do. And I am doing this here based off of its result in the 3D racing game, which was really quite well done. So I am hoping that this has the capacity to go ahead and do this.
All right, let's check out our PC simulator. All right, unfortunately I'm not actually seeing anything drawn here, which is quite upsetting. Supposedly, we have a fixed version right here, which so unfortunately it doesn't really seem like we're going to get a successful result for this, but this is a new high-end test that I have been waiting to test on models of the next generation. So, unfortunately, I suppose that is going to wrap it up, but I will say I was rather impressed with the performance of this model. Really based off of the one 3D racing game result I saw, that was by far the best result in that specific test that I've ever seen. And that is a more difficult test and meaning a lot of models I've tried it with have completely failed or done varying levels of non-working result. So, it was really good to see this. There is the possibility that the OpenRouter configuration may still be a little messed up as was reported a few hours ago by some folks who used it, but unfortunately this was really the only real way I could get this working in a more traditional chat response style test. And I'm going to say based off of some of the responses I received here, I don't think there's an issue now with the configuration. And if there is, then the model is going to be even more impressive. So, I suppose that's a good thing either way.
Really, that is probably going to wrap up what is an initial test of Minimax M2. I'm going to say first impression, I am impressed. And especially again considering that this is something that can hypothetically be downloaded and run locally. I think this may end up being a very strong contender in terms of low price, uh, fast, and cost-effective, low price agent coding uses. So, that's going to conclude today's video and test of Minimax M2. If you have any questions, please feel free to leave them in the comments.