Kimi K3 amb OpenCode: per què funciona millor com a subagent que com a orquestrador

Kimmy K3 just came out and this is really a historic launch for open-source AI models because this is an open-source model that is on par with Fable 5 and Chat GBT 5.6 Soul across many different benchmarks. According to Kimmy's test that they ran themselves, we see K3 right up there with Fable 5 and 5.6 Soul across all these different benchmarks, even ranking number one on S. E marathon, which is a benchmark that takes 20 realistic multi-hour software engineering tasks and uses those to challenge these AI models. For K3, we're looking at a 2.8 trillion parameter model, which is really impressive. It blows all these other models out there out of the water compared to Deep Seek V4 at 1.6 trillion and Kimmy K3 at 2.8 trillion. Absolutely massive model. K3 has really strong long horizon coding performance for game development and digital creation is extremely powerful as well. It can turn images and videos into fully playable interactive experiences. some of the examples that

it shows us here along this carousel. It shows a 3D openw world game here, a 3D um mind sweeper game right here, and a cyberpunk web swinging game, which is really impressive that it pulled this off. That's pretty insane, actually. And then for knowledge work, it's very good at creating these different kinds of interactive reports for different subjects. They take a little while to load on the on the page, but they're pretty interesting to look at. Also, for video editing, I thought this was very interesting how it said that they used K3 to edit its own teaser video from 56 source clips. Now, as nice as the official release report is, I like to look at this site here called artificial analysis because they perform their own runs of the models. And so, it's a little bit more trustworthy because there's less likelihood that there will be some bias skewing the results, right? So, when we look at what they found in their reports, they measure it like this across intelligence, speed, and cost per

task. intelligence. They put K3 right behind GPT 5.6 Soul on max reasoning effort. And they actually put K3 above Fable 5 and 5.6 Soul. So, it's a little bit faster than both of these models, which is really impressive. And that's the output tokens per second. But one of the problems with K3 that I've seen as I've been playing around with it inside of Open Code is that the user experience really isn't that great. It even says it in their official blog post that it's not as great as it is with Fable 5 or 5.6 Soul. And I feel like it really spends a lot of time thinking about things and it does some weird things like for instance sometimes it will uh think the code out like it will have these huge code blocks. It's just in its thinking so it's not writing it. I think you lose some of the cost savings there and some of the thoughts are kind of weird as well. But when we look at where this model is sitting compared to all the other ones that have open weights, I mean, it's just a huge jump that it's in the same like area as 5.6 Soul and Fable

5. Like that's really amazing. So, we see how these open source models are really getting better and they're actually starting to be able to kind of compete with these more proprietary models. And on artificial analysis, it shows us how it came to these conclusions. It shows us the individual benchmarks that it ran and where each of these models ranked on each of the benchmarks. When we compare Kimmy K3 with Fable 5 and 5.6 Soul on Open Router, we can see the difference really clearly according to the price. For Fable 5, we're looking at $10 per million tokens in, $50 per million tokens out. For 5.6 Soul, $5 in, $30 out. And then for K3, $3 in and $15 out. And the only difference as far as the modalities here is that Kimmy K3 doesn't support the file input such as um 5.6 Soul does and Fable 5 does. And for right now, there's only Max Thinking Effort that's available. And the full model weights aren't going to be released until July 27th, 2026. So that

means all of your requests are ultimately even if they're going through open router, they're ultimately going to be run by Kimmy directly. So that means that you'll be subject to whatever their sort of policies are regarding data training and etc. So if that's something that concerns you about running a model that's not being hosted on servers within your location, well then maybe you would just want to wait until July 27th. This is another really interesting insight. According to arena.ai AI K3 is ranked number one when it comes to front-end code. So K3 is very capable of making high quality front-end designs. So the first task that I had it do was to create four different versions of a productivity app. And here's each of those versions. We have this one. Uh this one which is more like um a modern tech grid, you know, with the glowiness and everything there. And then this is more polished, like editorial, which this looks something like something you would typically get out of a Claude model. Like the design looks very similar to how Claude typically does things. And this one was kind of

interesting. This reminded me of like the Matteo's Salsa website. And then I had it make a marketing page for those four different designs. And this is what I got back, which I think it's really very uh a beautiful site. We got a carousel down here at the bottom. No issues. very stylized. These different widgets going on, very modern looking. We even have these different little like image pills going on between this text right here, which that's pretty interesting. And this little scroll effect, which is very nice. Testimonials, very reactive. And it even lets me go directly to that file here from the page if I wanted to do that. So, for all these tasks, I should say that I use the taste design skill and I use the superpowers plugin inside of Open Code. So those websites I just showed you, that costed 143,000 tokens and $2.87. So very cheap there. For the next thing I did, I asked it to make a miniame about building a fire before the sun sets. And this is what I got back. Dusk Trail. Everything is 3D. And if I turn that on, you can see that it even has

audio enabled as well, which we'll turn that back off. But I can move around. Everything is 3D. I can um I have a stamina bar. If I hit the shift key, it will let me collect things with E. I can drop a glow stick with G. And up at the top, I have a compass that shows me back where my campsite is, which it's somewhere. Ah, there it is. Right. And oh, I got some twigs there. Let me grab those. And as the sun starts to set, it will progressively get darker in the game. So, for this, we're looking at about 75,000 tokens and $1.32 spent. So, I would say that was a success. Now, this is the hardest thing that I asked the model to do. I have this very small Django app that I use to track my workouts on my phone. I asked the model to refactor the code from Django to Nex.js. And it ran for a while. It spent about 193,000 tokens and $4.92, but it really took a very long time. It only got to task two out of 17 and I

just felt like it was more trouble than it was worth. I think the way to go is if you use a model with a better user experience and that has stronger reasoning capabilities such as GBT 5.6 sold or Fable 5 and then you were to use Kimmy K3 as a sub agent. And you can do that inside of Open Code by modifying your Open Code.json. Like here I created this sub agent and I called it Kimmy K3. And it's really simple. You just put the model that you want right here. So, open router moonshot AI Kim K3. And you can modify the prompt along with any sort of permissions and skills that you'd like there. And so, as it was running through the superpower skill, I just told it to use this specific model. And so, we had GBD 5.6 soul running as the orchestrator model and then it would spin up fresh Kimmy K3 sub agents to actually do the work. And I think that was actually a lot better. So for this entire experiment, I wanted us to see how much we could get from the model with just $20. So you can see I hit about $19.54. And with this method, we were able to get through the first four tasks. So

that's the brainstorming, that's the spec, that's the plan, and then that's the actual task execution along with the quality review gates. We got to task number five, and I saw that I was getting close to hitting the limit. So I said for these next tasks, we just need it quick. So no gates or test. And so from there, it was about like at $14 we had spent. So I had about $6 left. So with that last $6, it got to task number nine, it looks like. But what's really impressive is if we look at these sub agents to see just how long some of these were working. Like this one here, 57 tool calls, it was running for 29 minutes. 29 tool calls, 17 minutes. Like that's a really impressive run. Like here's another one. 57 tool calls in 29 minutes. So we have a lot of runs just like that. So it's definitely works very good as a sub agent. But here on the previous run, I'll show you some of the weird stuff that it was doing whenever it was trying to migrate that Django database. Like this is what I mean. It was just um generating a lot of like code in its thoughts. And I don't think I've ever seen a model uh do that before like so much like this. So, this is part

of the reason why I decided that we needed to choose a different strategy. You can see what sub aents you have inside of open code just by doing the at key and then it will pull up a list for you there. If you know for sure that you want to modify the default model that one of your built-in sub aents uses, such as your explore agent or your general agent, then you can just overwrite that in your config as well. You don't have to make this separate sub agent like I did here. I just did that just to make it clear about what's going on. In this video, I said that I use the Superpowers plugin for all of these different experiments. And so, if you've never heard of what that is, it's just a complete software development methodology for your coding agent. It's built on top of these different composable skills. And it's a very helpful method of working with coding agents. I have a video on my channel if you want to go deeper into that. And if you want to keep growing in your AI skills, then you should check out my school community, the AI Cheddar Chase. It's a place for us all to continue to learn AI tools together and business strategies. So, if you want to check that out, the link will be in the description. And you can check out that superpowers video I was talking about right over here.