China’s Kimi K3 DESTROYS GPT 5.6?

Kimi K3 from China just dropped. We already tested it. It literally just dropped about an hour ago and it is super impressive in terms of what it can build, how powerful it is. It's also really good at building 3D games, as you can see.

And today I'm gonna be showing you exactly how to use it, how it compares versus Fable and Soul on the benchmarks, and also why this is such an interesting model. So let's get straight into this. If you're not familiar with Kimi, this comes from Moonshot. This is a Chinese AI company.

And previously, Kimi K2.7 came out. That wasn't too long ago. Obviously, it was never quite Opus 4.8 level, but this new release of Kimi K3 is a huge step up. I'm actually testing it right now on a bunch of different tests, but it's so new that I haven't had time to test it out and everything.

However, what I can say is if you look at the benchmarks right here, you see that Kimi K3 is actually way ahead of Opus 4.8. So it's actually surpassed Opus 4.8. It's right up there with Fable 5. Obviously, Fable 5 and GPT 5.6, just beating it on the SWE, but K3 is right up there with 5.5 and beats Opus 4.8.

Let's have a look at Terminal Bench 2.1. Terminal Bench 2.1, Kimi K3 is right up there with GPT 5.6 Soul, and it surpassed Fable 5 and Opus 4.8. This is according to Leo, who tweeted these benchmarks out earlier. Today, thanks to Leo, and you see here Frontier SWE 81.2 versus Fable 5's 86.6.

So obviously, Fable 5 is still beating it, but on many of the benchmarks, it's either beating GPT 5.6 Soul, or it's beating Fable 5, one or the other. Now, we've actually already built it into our agent operating system. So the great thing about Kimi, and it's very unique in this way, is that it's pretty cheap to get on the coding plan. Then you can plug it in the CLI, and then you can build whatever you want with it.

So we have Kimi K3 already plugged into our agent operating system, which means it has access to our memory and all the other systems and workflows that we have inside here. We have a workspace and everything else, and this is a pretty powerful model. So you're basically getting frontier numbers from what should be an open source project. I would expect Kimi K3 to go open source.

Kimi K2.7 itself was open source. So usually the Kimi models are open source. Now this is a 2.5 trillion parameters model with a 1 million token context window. And for comparison here, Kimi K2.6 held 256 K tokens.

So K3 literally quadruples the previous models. You might also say like another Chinese model launch, should I care? I would say the ones to pay attention to these days are GLM and also Kimi. And these are two of the best Chinese AI open source companies in the world.

And this is getting seriously impressive now. It's getting to the point now where Kimi K3 is absolutely crushing Opus 4.8 on the benchmarks. So it's a huge step up, very impressive. We've tested it, it looks pretty cool.

The benchmarks look amazing on it. Arena AI actually tweeted about it as well. So it's now number one in the front end code arena, surpassing Claude Fable 5. That is a 17 place jump from Kimi K2.6.

So Kimi K2.6 was at number 18. Now it's ranking number one on design arena. This is a huge step up and this is massive news. And by the way, that was only tweeted 17 minutes ago.

So this is really breaking, but I think it's gonna be trending like crazy pretty soon. You can also see here BridgeMind. BridgeMind tweeted that Kimi K3 just beat Fable 5 on his BridgeBench Horror House game test. And it looks absolutely amazing.

I mean, that's what I've seen as well. The 3D builds you can create with Kimi K3 are awesome. We also tested it for website design. It still feels a little bit like generic AI fluff.

So it's not all good. Just wanna show you the benefits and also the disadvantages of this. When I've tested it on website creation, not that nice. So I would like to plug in my web design skills into this first and then try generating an output with Kimi K2 point, sorry, Kimi K3.

And also what I'd be interested to test here is Hermes Agent with Kimi K3 alongside it. I think that would be a lot of fun to test. So we've actually done that over here. So we created a new profile with Kimi K3 plugged into Hermes.

Now bear in mind that Claude, you can't plug into a agent like Hermes Agent on the subscription. But if you're using Kimi K3 and you're on the coding plan, then you can use Kimi K3 with Hermes Agent and you can use it agentically. And it's pretty fast to reply. Seems to work pretty nicely.

Feels really smooth right there. Now here's another example of a website that we actually built using the web design skill. And that actually looked 10 times better. So what I would recommend with Kimi K3, I think like straight out of the box, you're still gonna get your sort of generic AI fluff websites.

But if you look at this website that we've built, the full blog looks really nice. And that was fully generated with Kimi K3. Pretty awesome. Now, if you wanna see something pretty funny and this shows you how bad in comparison Kimi K2.7 actually was.

So we've got this on GoldieBench. This is the old Neon Racer that we tried to create using Kimi K2.7. And just to show you what a step up this is, this is the new version, the racing game. And it looks absolutely insane in comparison, right?

The difference in quality between K2.7 and K3 is unbelievable at this point. So I'm really excited to test it out more. The website looks great as well that we generated there. And I'm excited to test out Kimi K3 more inside the CLI here as well.

So exciting times, big step up. We're gonna be testing out on all 50 GoldieBench benchmarks on GoldieBench.com, which is gonna be exciting. And some people think like, you know, using Frontier AI models means paying Frontier prices, but actually if you're using something like K3, you get a Frontier model, but it's way, way cheaper. And the coding plans are way cheaper as well.

You also might say, okay, Chinese open source models, usually not very good. If you've looked at the recent run of GLM 5.2, this stuff is just getting better and better, right? Like I'm really, really fast. They're always releasing new models each month and it's scary how fast open source models are moving.

So thanks so much for watching. If you want to get our full system with Kimi K3 plugged into the agent operating system, we have Hermes agent, we've got all of these amazing workflows with Hermes agent, even outreach tool for using this sort of stuff. And you've got the Kimi code system plugged over here. We also have GLM 5.2 over here.

So if you want to get that, it's all inside the AI Profitable Boarding, link in the comments description or go to the AI Profitable Boarding.com. This is my AI automation community for saving time and scaling with AI automation. Inside the classroom, if you want our agent OS system, you can get that over here. And we give you a video tutorial, full step-by-step guide, and a zip file to install.

And then we add new daily tutorials, as you can see, so you can get the most out of this. We also reviewed Tencent HY3, which is another Chinese model that dropped last week. Not quite as good as K3, I will say honestly. Inside the community, you can ask questions, get help and support.

I personally create a video tutorial for members based on their questions every single day. And then inside the calendar, you can jump on weekly coaching calls, get help and support in real time. Inside the map, you can meet people in your local area who are building with AI agents, like Kimi K3. And that's all available, link in the comments description or just go to the AI Profitable Boarding.com.

Thanks for watching.