Kimi K2.7 + Hermes Agent is INSANE!

So Kimi K 2.7 just dropped. We've already plugged it into Hermes agent, but not just that. I've actually done some pretty amazing stuff with it. Let me show you this, for example.

So what we have here is a full video. What if you could describe an app fully generated with a team of AI agents using Hermes and Kimi K 2.7. Let me show you what that looks like. One sentence and have it built and bug fixed before you finish your coffee.

And having, what if you could describe an app in one sentence and have it built running and bug fixed before you finish your coffee, that's Kimi K 2.7 code, Moonshot's strongest coding model, not a chatbot, a builder, 256,000 tokens of context. It writes whole apps from a single sentence, runs them, finds the bugs, fixes them again and again until it works. I plugged it into my Hermes agent crew on a flat monthly plan, just a frontier coder sitting in my team, building real things while I sleep, no handholding, no back and forth, just working code. You can do the same, stop renting brainpower by the token, own the builder, Kimi K 2.7 code is live in.

So just to be very clear here, the whole AI avatar is generated with Hermes and a combination of HNC light, the editing and the design was done with Hermes and Kimi K 2.7. Then we also have the music and the voiceover that was done with Kimi K 2.7. Plus you've got all these portfolio things that we actually created using Kimi K 2.7. And that was actually plugged into the system as well, right?

So every single element of what you're watching, even like, for example, the camera zooms was done with this team of Hermes agents. Now, if you want to see what that looks like in practice, essentially we created a team of AI agent profiles inside Hermes with Kimi K 2.7 plugged in. And you can see, for example, we have one there. Now, what we can also do from here is we can actually speak to Kimi K 2.7 inside the chat.

And with the music, you might be wondering, how was that generated? So we actually have a music agent with Hermes plugged in, and that's basically a combination for Kimi K 2.7 as a brain, Hermes agent operating it, and then as the API for generating the music too. Now, if you're wondering how was the whole video created with a team of orchestrated agents? Basically what we did is we plugged in a task into our Kanban board with a list of our agents.

So for example, Kimi K 2.7 there and Hermes agent. And then we plugged it in here, triaged the task, assigned it to different agents. So for example, we've got a producer that generated with Heija the avatar video with the face and the voice. Then we also had an editor that edited the video that was generated.

And we had, and this is very, very important, we had a judge that was looking at the video, watching it with Kimi K 2.7, scoring the video, and then iterating it until it was actually amazing. So you can see eventually, first of all, it scored six and then seven, then nine. And then it kept fixing the bugs and eventually it proved it a nine out of 10. Right?

It didn't give it a 10 out of 10, it was a nine out of 10. And then you can see here, for example, we've got a script writing AI agent with Kimi K 2.7 as well. And we also have the iterations in terms of how this judge scored it. So it scored it on these different dimensions inside our AI agent.

And so everything that was created here with that video was done with Hermes agent and a team being automatically orchestrated with the Kanban board and Kimi K 2.7. So if you haven't tested out this model, you might think, oh, it's just another model that's dropped. Actually, it's really powerful. And one of the great things about this is it's kind of like, it feels almost like Claude when you're coding with it, you know, and I'll show you some examples of what we built with this.

So if we open this up and this wasn't created in one, you know, one, uh, shot, but it was created with Kimi K 2.7. It's like a 3d game. That's pretty cool. And the cool thing about this is the, when you're using this, you're using the coding plan, right?

So it's kind of like you pay for the subscription and then you can plug that into Hermes agent and then you get a lot more tokens. And so you can build multiple things or you can do long horizon tasks with Kimi K 2.7. And it's pretty chill when you're building with it. And also you can work autonomously in the background.

So for example, when it was actually creating some of this stuff and I'll show you another example in a second, when it was creating some of this stuff, like actually created a operating system. When it did that, it actually judged its own work and then went back and improved it based on what it saw. So it didn't just accept what we did. So this was not with Hermes agent, but it shows you the power of long horizon tasks with Kimi K 2.7, which you can do with Hermes agent.

So for example, here, if we scroll down and then we go to OS, this is an operating system that Kimi created with different apps inside the operating system. So we have the about section here. We have the notes tab, which we can save notes to, close it, then open it back up with the note saved, which is pretty crazy. It created like a calculator app, a paint app where we can draw with it as well.

And this was all a clock as well, which is actually telling the right time. And then you can have this inside your operating system with multiple apps. Right. How mind blowing is that?

And you can do this all with Kimi K 2.7. So if you're thinking, oh, this is just another coding tool. Honestly, it's pretty powerful from what I've tested out as well. It also created some other cool, fun stuff, as you can see right here.

But I think it's super powerful because, you know, a lot of people are looking for like a good agentic brain to give to their Hermes agents. Hermes is powerful, but on its own, without a good API, it's not so good. And also, you know, see you want to use something that's Claude level. But if you're using Claude on the API, well, it would get too expensive.

If you're using Claude on the subscription, well, you can't actually plug that into Hermes. Right. And so this might be a viable alternative that is similar to Claude level and still super powerful. And again, like you can manage multiple agents here.

Now you might want to do as well. How is it for just using Hermes inside the chat? So here's an example. We've got KimmyK2.7 plugged in as a separate agent profile.

And if we ask it, okay, I don't know, like open up juliangoldy.com for me. We'll plug that in. Let's see how long it takes. And I did notice it was a little bit slow when it was going back and forth.

And I don't know if that's because it's a reasoning model or if there's a way we can speed it up, but if we ask it to do a task like this, it should get it done. Let's see what it does. I mean, when I was just asking, like, are you here? It still took like seven seconds to reply.

So it's a little bit slow, but if you're running long horizon tasks, then this is powerful. It is powerful. Um, also another cool thing that it did is it like created this showreel as you can see. So it created this full video with a full like showreel of everything that it created inside the video itself.

Now this is using a combination of hyperframes and KimmyK2.7, but you can see the web OS that it created here and create some other cool stuff too. So again, like it is pretty slow. If you get it to do tasks, like just doing computer use or browser use, it is slow. So I'm not going to tell you it's perfect, but you can see that it completed the task right here and it read the whole page.

So I know that it's actually navigated to that page because it shows, for example, like testimonials that we have from real people right there, which is pretty cool. As it can open tasks, it can work as an autonomous team. It can create awesome videos as you saw. It can create like full portfolios of different stuff.

It can create games as well. And really like, it's just, you know, you chat to it and then it's good to go. Uh, these are some other things that we built with it. So if we open this up, you can see like pretty cool animation and then I can move my mouse and it scrolls around and changes the screen based on how I scroll.

Here's another one, right? So it follows my mouse around over the screen too. But I think for me, like the most mind blowing thing was, cause I get this question every time, like pretty much every day inside the AI Profitable Like how do you create videos that are actually good and nicely edited, etc. Using AI.

And I think this might be the best way that I've seen so far. Again, it took like about 90 minutes for our team to create that full edited video. But still, was it good? Yes.

Was it useful? Absolutely. Was it better than some of the stuff I've seen from my own team of editors? Absolutely.

And you might agree as well. Felt very realistic when I was watching that video, I didn't feel like robotic or anything like that, which is pretty crazy in itself. And so it can write the whole thing. It can code for you.

It can check its own work. It can judge its own work. It can plan stuff. Um, yeah, pretty powerful model.

You might be wondering, okay, what are the benchmarks like? Honestly, do I pay much attention to benchmarks? Not that much. Let me just double check this stuff.

So if we have a look, Kimi K 2.7 code. Yeah. So it's 262 K context. Let's have a look what apps it's being used with the most.

So Kilo, Hermes, Klein, OpenCore, Pi. These are the top five apps being used with Kimi code right here. So yeah, it can do this stuff and we can chat with it. I do like the fact that you've got now agent profiles and you can just chat with you directly.

You might say as well. Okay. What is this UI that we're using? We're using the agent operating system, which is inside the air profit boarding link at the comments description or go to the airprofitboard.com.

If you just want to just go straight into the terminal and change your model, you would just go to Hermes model, type that inside your terminal, and then to switch to Kimi, you go to here, right? And then you want to select Kimi coding plan. Otherwise you're going to be using API, which could be more expensive. So that's the way I recommend using it, how to set up, et cetera.

And then, yeah, we've built in the Kimi code CLI directly into the agent operating system as well. So then you kind of get the best, best of both worlds. So you need something like a big coding project. You can go inside here and you can see everything that you've created inside the workspace.

Right. If for example, you want to use Hermes agent, then you've got Kimi K 2.7 there. You could also talk to it. You could use it inside Hermes Jarvis, which we've created to voice activate this.

You could have the studio there and everything else that we've built as well. And I think probably the true power of this that's really unlocked is having that system where you've got the video agents working together. This was another Kanban board, as you can see right here, where it took the information and then plugged it in. But again, amazing stuff.

Amazing. Now you might be wondering, okay, what the other benchmarks like? So it's a mixture of experts model, 1 trillion parameters, 256K context window built specifically for long agentic coding tasks. These are Moonshot's own published benchmarks.

So it shows a big jump on Kimi K 2.6 previously. I haven't seen any independent benchmarks yet, but still. Is it good from the outputs? Yes.

So how does it compare versus Kimi K 2.6? Well, it went from 50.9 to 62.0 with Kimi Code Bench. For Program Bench went from 48.3 to 53.6. MLS Bench Lite went from 26.7 to 35.1.

Kimi Claw, which is interesting. If you're not familiar with Kimi Claw, if you go to Kimi.com, basically what this means is like you can use Kimi Claw, which is kind of like a cloud hosted version of OpenClaw and you can use it in the cloud here. So if you like non-technical, it's a good way to use an AI agent. And so you can see the benchmarks here went from 42.9 to 46.9.

So like over a 10% improvement. MCP Atlas. So can it use tool use? Yes.

69.4 to 76. MCP Mark Verified went from 72.8 to 1.1, which is pretty good. Head to head so far versus Opus 4.8. Again, are these independent benchmarks?

No, but on MCP Mark, an agentic tool use benchmark, K 2.7 code scores 81.1% ahead of Clawed Opus 4.8 at 76.4%. So it is crushing Clawed Opus 4.8 on benchmarks. And if you look at the quote on this release, K 2.7 code uses 30% fewer thinking tokens than K 2.6 while scoring higher across Moonshot's coding benchmarks. So it uses fewer tokens, but it also gets more done and does it better, which is pretty useful, pretty cool stuff.

You might say, well, isn't this technical to like set up all this stuff? No, like loads of people inside the AI Profitable Automate are getting awesome results with this stuff. So if they can do it and I can do it, you can do it. I'm not a coder, non-technical, et cetera, but we can all use this sort of agent operating system.

We can all use Hermes. There's no barrier at this point for being technical. So that's basically it for the system. How to use it, how to build with it, how it works, et cetera.

Pretty cool. If you want the full agent operating system from me, you can grab it inside the AI Profitable Automate, it turns Kimi, Clawed, OpenClaw and Hermes into one system with shared memory. One dashboard is built by describing. Every new agent you add makes the whole thing more powerful.

And the Kimi setup, every prompt, the Obsidian memory setup, coaching calls, et cetera, are inside the AI Profitable Automate. So if you go to the classroom here and then you go to the agent OS section, you can see it was just updated today. So you get the video tutorial. You can see the last update date.

You can grab the zip file for installing the whole system. And then you also get my best automation blueprints. So for example, Hermes just released the automation templates update yesterday. We've already got a tutorial on it.

Clawed Fable 5, Hermes Idea Factory. Clawed Fable 5 is already just being removed, but you get the point. Like as soon as this stuff comes out, we release new updates and tutorials on it. Inside the community, you can ask questions, get help and support, et cetera.

And I personally answered this stuff too. So for example, you know, if someone asks a question within hours, we usually reply to them. Like she never asked a question here about our agent operating system that she built using our methods. And it looks pretty awesome as you can see right here.

And so we've already replied to her, right? And that's how quickly we move inside here. You can also jump on weekly coaching calls, get help and support whenever you want to. And then also there's a map where you can meet people in your local area who are using AI agents like you, which is great because we've got people all over the world building with this stuff and you can easily connect with them using this system right here.

And that's all inside the AI automation community that we have. And it's just focused on helping people learn and grow and build amazing stuff. Again, like let's just review what this has done. Cause it blows my mind.

We created this amazing game with it, which is like a 3d game. Right. Pretty cool. We also created a fully edited video that's beautifully designed.

As you can see right here, with an app in one sentence and have it built, running and bug fixed. We created a web OS and we also created like a Skyrim style game as well, as you can see right here. So it's kind of like a, an RPG style 3d game where you can navigate the landscape and go off and do cool stuff. Right.

And you can build this with KimmyK207. Bear in mind, like it just dropped a few hours ago. So, you know, within a few hours, we've just been building this, not like it's taken a lot of time. So hope to see you inside the AI Profitable Boardroom.

Thanks for watching. I'll see you on the next one. Cheers. Bye bye.