China’s Kimi K2.7 is ABSURD!

**Julian Goldie** (0:00)
China's Kimi K27 is scary good. So I want to show you exactly why that is. Three days ago, Kimi just dropped from a really small lab in China called Moonshot AI, and it's building some absolutely amazing stuff. So it came out on June the 12th, and it costs almost nothing to run.
And this thing just came from a tiny lab. It's open for anyone to use, so it is open source, which means that you can start running it for free. It's already available on Olamma. I want to show you what it does, what it can't do, and what this means for you and your business, because there's a way to use this right now. Most people are going to miss. And let me show you the before first, right? So the old ways you get an idea, you think about building it and it just kind of disappears, or you build it and you never find it again. What this is doing is building some amazing stuff, and it was super simple. So, for example, here we have a full agent operating system. This is literally an operating system with different apps. So if we type something into that, it's just going to automatically save. We can close it, open up again, and it creates this literally an operating system. It's like building windows, but we just did it with Kimi K27. Obviously not as complex, but you get the point.
Here's another example.
It can create some pretty cool animations and stuff. We actually created a full game, as you can see right here. Here's another game that we created, and we even created like an open world game, as you can see right here, right? And so you can build some absolutely amazing stuff. This is pretty wild how good it is, and you can do it in one sentence. Now, if we have a look at the benchmarks here, we can see Kimi K27.
Measured is a one trillion parameter mixture of experts model, only 32 billion active parameter per token, with a 256K context window. It's pretty small actually, but it's built for long, agentic coding tasks, which means like when we plugged it into the agent operating system, it could build some more stuff, simply because we could have it coding literally like an hour, just on its own, just getting stuff done. For example, it coded this out, everything that you see in here, it literally coded out itself, which is pretty insane when you think about it. And it did this all in like, I would say a couple of hours. You got all of this done in parallel. Even for example, created a full video that's fully edited with an AI avatar. And you can see that right here. And one of the biggest differences between, for example, something like Claude and Kimi K27, is that this has a coding plan, so you can actually plug it directly into Minas. By the way, that is not me. The voice on that, not me. The editing on that, not me, right? The scripts on that, wasn't me. I literally just gave Kimi K27 a prompt in the skill to create videos and Boom Shackalaka could just go off and create all this awesome stuff. And then we've got it plugged into our CLI over here, and we already have it inside Hermes. So this is a really powerful model, pretty cheap to get. I think I'm on like the $99 a month plan and have come nowhere near using all the tokens. You can see how it performs on benchmarks and it's pretty powerful stuff. GLM 5.2 just came out recently. Which one do I prefer? Bear in mind, they're both from China. I actually prefer using GLM 5.2 for coding in general. I think it's more, it feels more like Claude to be honest with you. But when it comes to Gentic stuff, I found the Kimi K27 replies a lot faster when it comes to using something like Hermes and that sort of thing. So that's just one thing to bear in mind as well.
But test out for yourself, see what you think.
And so nowadays, you can just describe what you want, you get it built, it checks itself, and then you get the garrily. And this is something else that we did recently, is that when we're using something like Kimi K25, we have different Kanban boards for different projects. For example, like this one for building out a dragon game. And then we have different profiles for creating each part of the report. And not only that, if you want good quality stuff, and I recommend having this in place, is like have a Kanban board or a team of agents, and then make sure that you actually have someone who checks the quality of the work, and if it's not good enough, they send it back. So here's an example. The video judge right here, they iterate until the video itself is scored by the judge as above 8 out of 10 And if it doesn't hit that, we're not going to use it. And it just keeps going back round and round and round until we eventually get the video that we want to view. And so that was how we got that video before created. Again, you can see on the Kanban board here in terms of how we created it. And it's pretty nice on the editing. It's probably better than most of my team could do at this point, to be honest with you. And this was fully automated with Kimi K27. So if you've not tested it, definitely recommend that pretty powerful stuff.