Qwythos 9B is INSANE (FREE + Local + Open Source)!

**Julian Goldie** (0:00)
We're going to be looking at Qwythos 9B, which is a new local model, as you can see right here. It's come directly from, it's available on Ollama to host, and there's many different models. You'll see the context window here is one million token context. I'll come on to that in a second, but essentially this is designed to be like a Claude style creative reasoning model with a Qwen 3.5 9B base. And so what you can do here is you can easily install it from Ollama. If you don't have Ollama set up already, this is a way to run three models. You can use this terminal command. Once you've done that, make sure you have Ollama running. Then you can run this terminal command and you can download the model and start running it directly. We've already plugged it into our Agent OS system so that we can test it out. You see, for example, we've got it inside the build section here. I'll show you what was built within a second, but we can see all of our creations inside the local engine of our agentic OS. So I'm going to run you through what it means, how it works, what it is, why it's different. And this is basically a free Claude style AI that runs on your own Mac. Now, is it like OPUS 4.8 level? Probably not, probably not. But is it decent and can actually create stuff locally? Absolutely. So Qwythos 9B is a Claude, Meafos style, creative and reasoning model that runs 100% on your machine, which means no cloud, no token cost, nothing leaving your computer. We've already wired it into the Agent OS local engine, tested it out and I'll show you exactly how it works. So you can see the details here. So it's pretty small as a model, only 5.6 gigabytes when we ran this version and it was free to get from Ollama.
You'll also see some stuff that we created with this just for fun, you know, to test out, see how it works. So we create like a little to do list here. As you can see, we can add new tasks. We can clear everything that we've created. So you can create like little mini apps. This was like a digital clock that we created that runs locally. And this is like a little snake game. Classic, classic. So we put it to work. We tested out. As she works, which is pretty cool, and it's running privately. Here's a few more things. So actually one thing that surprised me genuinely was that it could create like a fairly nice design. I've seen I've seen like frontier models create worse stuff than this for the landing page, which is pretty nice. It created some stuff that was a little bit glitchy, but even create like this calculator that seems to work as well. So we tested it with like six apps, level three, running it with Ollama. You can see how it works inside the terminal right here as well. And there's two models I've seen recently that have been interesting for local, which is Qwythos and also Ornif 9B as well, which is something we tested out. From what I can see so far, I would genuinely say that probably Ornif 1 is outperforming it. I've got another tutorial on that, but so far it's pretty good. If you're wondering, OK, what's the build speed like? So you can see the Qwythos was a lot slower. Sorry, it had more tokens per second.
On short and build designs, it's a lot smaller as a task. It can build working apps. And the context window is insane for this, which is one million tokens.
And also, you know, Ornif and Qwythos are both local free models. Now, both of them are great, you know, but for the same work, Qwythos actually came back twice as fast, which is pretty cool. Now, you might be wondering, okay, why is it called that? So it's a stack of four deliberate moves laid on top of an open Qwen 3.5 9B base. That's why a model is small enough to run on a laptop. I mean, five gigabytes is really, really small for a local model. It can write a reason like something far bigger. So the way this works is like, it's based on Qwen 3.5 9B, which is a strong open 9 billion parameter base model. Then you've got something that's trained on like Claude style reasoning and creative. From there, it has a million token context window. So the context window has been stretched and it's also quantized with Ollama and Llama CPP.
So it can think like Claude locally, it's designed for creative stuff, big context window, reasons and cause tools, which is good for agentic. You can use it like so. You can also see the benchmarks here as well.