North Mini Code de Cohere: prova local d’un agent de codi

This is oddly oddly similar to a front end I got that was created by Fable 5. I immediately noticed a similarity between the two. Today we're going to be taking a look at a very exciting release from Co here and this is called North Minicote. Now this is in Apache 2.0 licensed 30 billion per hour mixture of experts coding model and it is from a lab that entirely has created it themselves. So oftentimes at least personally, when I see a new model that is small and able to be run locally, one of the things I always want to check is if it is just based off of an existing base model, not that there's anything wrong with that, but I find models that are from the ground up completely their own. And sometimes have like fun vibes to them and it's always enjoyable to test a truly new model that was created from someone. So Co here is based in Canada and as we can see right here, this isn't Apache 2.0 licensed model. They have been making cool local models or I should say open weights models for quite a while. So it's always nice to see something that is in a popular size come from a lab that's been around for quite a while. So before we get into it, please do feel free to subscribe as I do want that 100k plaque and let's talk a little bit about this model. So we see this is dated about 10 days ago and this has been out for a little while. However, I saw a post by the folks at Co here in a subreddit called local llama, which is centered around running AI locally and they had mentioned just today that this has a lot more availability in terms of folks being able to use it without necessarily needing to download and run it locally. One of those things is that it is now freely available on open router. So if we click in the search bar just from any open router page, it shows up as one of the top models here. And usually things that show up closest to the top when you search are either new or popular. So I would imagine a lot of folks will be seeing this. Additionally to that, this is a size in terms of this model where

it is fairly accessible. If you have a single 24 gigabyte GPU in your system, you will be able to run this at a decent quantization and a relatively nice speed. Now in terms of model specifics, there is of course a hugging face model card and this is a patchy 2.0, which is very good. It's just a nice free license, at least for open models to be released under a really anything I suppose. However, they do succinctly list some of that information right here. So the model size is 30 billion parameter mixture of experts with 3 billion active. It has a context length of 256k, which is not bad at all, and they do list the maximum generation of 64k tokens. That's not always prominently listed. So I like seeing that. Sometimes you have to do some digging to find what that figure is for certain models. And then optimized for code generation, a gented software engineering and terminal tasks, which is kind of good, at least in terms of the tests that we like to run on this channel. Now of course, there are benchmarks. And I do like to see that this is being compared with the model that will probably come to mind when most folks see a coding specific model of this size, which is the Q1 3.6 MOE at 35B. Well, it is not outperforming that. It does seem to stack up at least pretty decently in terms of some of the more coding specific benchmarks. They do also include the pool side models, which I have not yet tested on the channel, but at some point I absolutely 100% will. Now in terms of how we're going to be testing this today, at least for the first couple of tests, I will likely just use this through open router. However, this is running locally on my RTX 6000 card. It is using an 8-bit quantization. And we see right here that this is specifically trained for compatibility with open code, which I do very much like to see because we will be testing this predominantly from open code. And it's being served just again to reiterate. It's an 8-bit quant of this model with full context on an RTX 6000 Pro Blackwall system. So for our first test, just through open router, we're going to be doing the Trident True Browser OS test. Now my

assessment would be just based off of the model size and seeing some benchmarks. I think it will probably produce a decent looking browser OS. I don't think the GTA clone will be functional. Just kind of a wild guess. So we can see one of the things here. Now this is being served by co-hear themselves. And the speed is very, very good. We see it's reasoning trace right here. The speed will translate over to when we're testing it from within open code as well, because it also does go very, very quickly from the Blackwell 6000 card. All right, let's check out our browser OS. Oh, okay. So we inevitably are going to have something that has unfortunately gone wrong here just in terms of the syntax. All right, I'm going to give this a follow up here because we unfortunately did get an issue and I'm still just wanting to stay in open router at least for the time being in follow-up tests. We're going to just use this through open code. So basically, I just gave it the information that like the background's not working. I gave it the specific error. I'm going to give it the option to fix it. However, in the meantime, I suppose I do have some level of interest in just checking to see some of these apps. Specifically, I'm just going to click on what I would assume to be the GTA game. All right, it did. I did not, I make the Chrome update thing. I'm about to spin out from that. I did not 100% have faith that it would actually create a 3G GTA game. Now, of course, the argument could be made that this is quite simple. However, for a model of the size that did it at this speed, this is respectable, especially because this is like, it's not bad. I mean, it's bad, but it's not bad. It's what I want to say. All right, that gives me some hope. We did have another game. Oh, maze. Oh, wow. Okay, maze the movement and maze perhaps should have been swapped with the GTA movement, but still, this is actually like, it's not half bad. No, don't even. Hello, help. That's actually cool. What? Open GTA. Oh, okay. Oh, what? How do you like put out this gore bridge and then put in a

sick sick special feature like that? I'm actually confused at this point. This thing seems to have some pep and what I mean by that is, okay. Let's, oh, flux climb. These are, um, like, I don't know, because flux is censored. So we're safe. Oh, wow. Okay. So the random photo I happen to select for the wallpaper is that, but this is definitely a significantly better looking browser OS now. We don't have a clock or anything. We have a fake file explorer. This is a decent good start. I'm going to say, I'm actually happily surprised just based off of the fact that the GTA clone, both games did open and work. My, oh, who's premature, because, okay, you know what? This is not bad. If we just like, look at this scene right here, maybe we'll block the weirdness out. That was just a random photo. And it even shows up in the taskbar. Pijon likes the model already. All right, theme light, happens if we swap to dark. Okay. Well, it, somewhat makes a difference, but not 100%. Now I did, of course, tell it that, like, the result was bonk effects. Okay. So the background didn't get fixed, but the issue we had in the first one before manually changing it where the text was appearing there and was weird. Did get fixed. And the GTA clones a little worse. So I'm going to go with the initial result for this, because the special feature with the voice activity was awesome. And then also we could upload our own background. So that fixed it for my purposes. So next up, we're going to be trying the beautiful static subway scene test. I am using this from with an open code now. I have put it into plan mode. And again, it's just being served locally from a 6,000 black well card. I also did change the background just on the browser OS right here to something that was a little more fitting. So it's as some follow-up questions that are good project location. Yes, current project directory. Creation tool. Okay, that I'm not 100% familiar with that specific question there. And I told it to use 3.js. Alright, I've noticed a few things here that are a little concerning. Mainly, just one ahead and

did all of this from within plan mode here in open code. Now it's at the point where it does seem like it may have produced something. So all right, let's just take a look at this and we'll see what we keep in mind that it did this from within open code. That's, you know, what? That's actually rather passable considering the size of the model. And the fact that this is an 8-bit quantization now everything we've done from within open code is just running locally on a machine. So okay, now frustratingly does our brightness slider work. Okay, it doesn't seem to, which is a little frustrating. And then of course we don't actually get the WASD movement. Let me check for any blatant errors. I'm going to just swap it back to build, not back. I'm going to swap it to build mode. Alright, so I've tried a few times to get this to fix the issue with the scene at least the syntax error that we had. Unfortunately, I'm not having any luck here and I swear these are like little people put in this station. Nonetheless, this does show some level of capability in terms of using 3.js and making 3.D scenes online, at least. Now out of pure curiosity, did the same test just through open router just to make sure that the quant I have in the model settings that I'm being run locally with are working. I definitely would say so because this is significantly worse than what we had generated from the local setup. So this makes me very bullish on testing this just through open code for the rest of the video. I want to try the 3D printer sim and I'm just starting this from within build mode because last time when I started it and plan mode it wrote the file anyway. So it's probably fine to just do this. Alright, that was extremely fast. It does say it has created this 3D printer sim so we'll take a peek and see what we get. I did put it in its own folder. Okay, well, where's the save it as an HTML file? Okay. Also, just out of curiosity because I'm noticing some of the open code behavior is not 100 percent as I would expect. I have ensured that I've run the specific instructions exactly

as they are listed right here in the hugging face model card where they do specifically tell you using it in open code using it with VLLM and then additionally the specific exact open code config do have set. So all of this is set exactly according to this specific model card. So I just want to reiterate that but it does seem like it has now produced the 3D printer sim. Very good. Now we can take a peek and we'll see what we get. Okay, that's a little frustrating because we don't actually get anything visual. However, this is probably fixable. Simple to fix syntax errors. So I'm just going to say fix this issue and we'll paste in the error and hopefully it will be a quick fix. So it puts sound in here and that's partially what's causing the problem is just the specific integration of that sound. That is where one of the syntax errors is this is what I meant in the beginning when it's interesting sometimes to test models that are completely different and not just based off of someone's existing base model because this behavior including sound in this result is not something I would have seen with a quen model of the size or the gem model of the size just including that in the printer result. So it's always interesting to test different models to see what sort of spice they add in on their own accord. All right, I just saw something where it said fixed and then I think I rewrote the entire file. Okay, I'm okay with that. I just, okay, that's unfortunately we had some regression in terms of the issues. All right, it's possible that this may be a little outside the scope of its capability here. And I don't imagine it should be though. It should be able to do this. I don't think this is going to work but I did just give it the self-contained C++ skateboarding test. It would be very very surprising where this to actually create a C++ file that was going to compile and give us any gameplay. I've just initiated it from within build mode and we're going to see what we get nonetheless. All right, it did write a file and it tried to compile it. It received some errors. So what it's doing now

is trying to fix those. We can see right here we got a bunch, but a few compilation errors. So it's now checking to see what other graphics libraries are available. And my assessment is it's probably going to rewrite the file now using a different graphics library. Still, that's actually not bad the fact that it went. It tried to compile the script that it wrote. It received an error. Now it's trying to fix that and writing a new file with perhaps a more compatible graphics library. So after a bunch of issues where it was compiling this C++ gate game and it was getting errors, it just did a compilation and there was no output, which is a good thing because it means that we didn't get any blatant errors. Okay, we have a compile game here. Unfortunately, it's not really doing much of anything to play. Compile. Okay, but you already did that. That's all right. Simple skateboard game console version press any key to start. Okay. It made it in the terminal using ASCII. That's just like I rotation. I have no idea what's going on here. I'm actually one to four strikes. Okay. Virokic 12 points and we see it actually took effect because our combo went up. I'm like bonus combo. I am dumbfounded by the problem-solving in creativity that was displayed by this coming to this specific point here where it actually created this game too run entirely from within the terminal. And the fact is it's actually interactive, but it's mainly text-based because the ASCII graphics don't necessarily- this is exactly what I mean when I say I like testing models that are entirely different from other models. Because they exhibit such odd behavior and interesting. Like I don't think this would have been seen with any model of this thing size basically the Q1 3.6 30B 35B MOE or the Gemma 26B MOE for Gemma this is art a restart and it did. Our combo got reset. Okay. W. I'm so enamored with this skate game. I'm going to push this much harder. So we're in the same open code session that created this quite frankly masterpiece and I don't say that's sarcastically. This terminal-based game was a masterpiece. I'm going to instruct it to now just turn this into a simple web-based game.

All right, it's not done, but it's put out two individuals skateboarding game files. Both are on the same size. I do just want to out of curiosity if nothing more take a look at what was created here. Okay. It genuinely transposed this terminal-only skate game to the web at least for the first one which is very interesting. It did a decent job getting the UI and everything. Let's just try this now. Okay. It basically did the same thing for both. I'd like to try a simple front end test so I've just said make me a beautiful website for a GPU rental business. Just start from within build mode through open code. All right. Let's check out our GPU rental front end side. Okay. This is not half bad. Power your AI, ML projects, good. We have 500 plus GPU models. I don't even know if that many exist. This is a nice little hover effect here. Let's scroll down. Good. We have some cards. As one would expect gaming and eSports FPGA integration. Simple pricing. Nice pricing card. These are the things I would have expected to see. So 3 GPUs for 49ies, then we have A100s and 3080s. Enterprise, white-use GPU rental. Okay. There's a broken image there which lets it down a bit. But overall this is actually a fairly competent result. And I will say the footer is nice and clean and well done. So I'm happy to see this. So it does have some front end shops we can deduce. I've gone back to open router for a moment and I've given this the flight combat simulator test. I just want to see what we get from open router. I almost feel like sometimes testing this in open code may work against it a bit because it just I've seen it come up with multiple different scripts like it will do two versions and stuff. It's just that some odd behavior with open code. Though I am using the complete recommended settings. So we'll just see what we get with this. So here's a flight combat simulator result. Okay. Start game. What the heck? This is this is just funky. I mean we do have some space to fire. Okay, there's definitely some activity to this. Okay, so those are health bars for the opponents

which is interesting. Let's try a role play test. I'm just going to do this through open router because it's easier to read the GUI. Hey there, I'm steeped. The guy who lives in sleeps with the toolbox will screws adapter and a key knife for what's slowing down your tech. We'll kind of laptop desktop or gadget as you stuck nowadays. Okay, these are all pretty simple cookie cutter. I'm not going to waste our time going through this word for word. Let's just see if it will have a little fun with us. What the heck? Okay, I genuinely normally they'd be like, oh okay, it really went into the role play, but from like a completely logical way. Like absolutely, like the IQ here was present. The EQ here was completely absent from this scenario. So it basically made up a PC repair job in Japan that went wrong the client. Small socket-based IT services firm location equipment, a fleet of thing-pat X1 carbon symptom laptop won't stay on for more than five minutes after startup. So basically we now have a table here of things that Steve went through when trying to fix these. My reputation took a small hit, especially after the cultural misstep of using the wrong power adapter. Then we had this extra details expand if you're curious section. Okay, and it's just talking about the power adapter issue that Steve had caused. I'm just going to do some more front-end tests because it's interesting if nothing more. I've asked it for a beautiful high tech website for Steve the PC repairman. I have to say this is this may be one of the audest models I've ever tested today. All right, I think this is our Steve the PC repairman website. Where did this come from? This is fantastic. Look, this is oddly oddly similar to a front-end I got for attacking with the diffusion gem model that was created by Fable 5. It had an eerily similar aesthetic to this. To that point, I do have that file present on the system as I would like to specifically reference what I'm referring to. So Fable 5 made this front-end right here and it's just odd that I immediately noticed a similarity between the two. So that's actually quite interesting. This is like a this is fantastic. This is absolutely spot on.

I'm not quite sure I don't have a specific piece of feedback for this model. It's totally been all over the place in this testing. However, this specific result right here it is possible that I'm extremely bothered up by the kind of synthwave aesthetic or like tronestatic, but look at even like the off-kilter hover effect in the get service. Now, core services we have nice hover effects on these. I have absolutely no idea where this just came from, but this may be so far the best result we've received. This is really like I really like this. That tech lane in Silicon Valley USA. I don't know why my dog won't go away now. The Ubuntu install in this system has just been a disaster from day one. But this is fantastic. I do want to try one more front-end test just because of how awesome the Steve the PC repairman was. So I've told it make a beautiful high-end high tech website for a custom PC manufacturer called PC System O1. There should be animations custom SVG art for the PC builds and other high tech things. So it may have to go ahead and create some little SVGs for the actual desktop builds. At least that's what I am envisioning in my head. So hopefully we do see that. All right. So here's our custom PC builder system. Now regrettably it was definitely influenced by the existing site that was there for the Steve the PC repairman site. So let's just focus on some of the custom assets that we told it it needs to have. Really though these like the cards, the hover effects on these, they're not that bad. Master builders. Okay. And then we have the same names as the fake customer testimonials here. But instead these are the master builders. These are some pretty aggressive hover effects. This is where I want okay build gallery. Okay. Unfortunately though these are some custom SVG assets. It's not necessarily, you know, this is pretty cool. I like this aesthetic and that may be influencing my decision to say this is cool. Okay that is a bit odd. But fortunately it only happens if you specifically hovered over this book builder. All right. Wow. It's significantly better than our GPU rental front end. I'll give it that. I'm trying one more front end test because the previous one

was unfortunately influenced by the result that was existing. So I've told it. Make me a high-end website for the Becky Jane Quantum Space. Do not search the system build only within this directory. Okay good. I had improperly assessed what I saw here. It was only looking at files in the current directory which there were non so this will not be influenced by any existing result that it's created on this system. All right. Let's take a look at the Becky Jane Quantum Space front end. Okay. Not bad. We do have some particle effects in the background. We have the BJQ so we're the Becky Jane Quantum Space. All right. Let's scroll down here. Quantum computing interesting effect on these cards and hover effects. Our services about Becky Jane Quantum Space. 50 patents filed contact us. Okay this site is not as cool as the PC system 01 which this was like taught here. Seriously this was a I mean maybe it wasn't but I very much liked it. So unfortunately doesn't seem to be properly loading at this point in time but that's okay good because we did see that this one was not as cool but this thing definitely has some interesting front end capabilities and partially what will lead us into the conclusion of this. I like testing models that are entirely creations of the lab itself and not a existing fine to not for base model. As I've probably said multiple times partially because we get to see interesting behavior like this. This has its own style. If you will now that looking back at it in retrospect the browser OS result that it did put out was actually kind of interesting because it did have that capability to open apps just through voice. So if we open this keeping in mind that the one big issue right here is the background isn't loading by default. So if we go into settings and then just choose a random image. Let's just here's the photo of myself for the car I guess just for the time being. So if we choose a random image we see the browser starts to come together significantly better. The GTA clone actually now that we look at it this isn't a half bad. Yes it doesn't really work and it's too

slow to move the vehicle but it did properly make both 3D games even if they weren't 100% playable and that was cool to see. Additionally this voice assistant here where we can actually launch apps was definitely not something I would have expected. So open GTA and then it does it so that's actually pretty darn cool and I did not expect that. So this seems to have somewhat of a pension for some form of web and front end design. I guess just based off of that and this really probably the flat out weirdest thing that I have seen ever is the C++ skate game result which and then we tried having it turned into a web game which unfortunately didn't work but this was a terminal based skateboard game and it did kind of work as we saw when we were playing it. The scores would go up and things like that. It was just such odd behavior that this is still probably one of the weirdest models I've ever tested and that's not necessarily a bad day. It's just very very different and very interesting. I would urge folks to try this if nothing more than like do a few front ends with it or something because you may see something you like that's very different. So overall that is going to conclude our first look in test of co-hears North Code mini 1.0. This was a very very very interesting model and a few folks had mentioned that they wanted to see the testing of this and I'm very glad that we did. So that's going to wrap it up. If you have any questions please feel free to leave them in the comments and thanks for watching.