A1 Agent: New FREE Chinese AI is INSANE!

**Julian Goldie** (0:00)
Today, we're going to be looking at a brand new model called Agent A1. And this is a new free Chinese model. You can run it locally. Let me show you what we've built with this so far. So we actually tested it out inside our Agent Factory, plugged into free Claude code with a local model. And it's pretty good. Like, I mean, check out this local landing page we created. Again, this is free. This is local. It can run privately. It can run offline. We could be on a plane. We can code with this bad boy. And it's pretty easy and simple to use and set up. So you can see some stuff that we built here. If we want to use it, basically, we can just go inside our agent factory here. And we could say, okay, build a keyboard, for example. And what I actually do is start using the live build and then start creating here. The other thing I noticed with A1 is that it's pretty fast to apply. If you're wondering how it performs on the benchmarks, we'll come on to that second. If you're wondering how it performs against all the other local models that I've tested recently, on the local leaderboard is now ranking number two. It's not as good as Qwable, but it has outperformed Gemma 4 on our test so far. And if you want to see some examples of what it can create, I actually created this open world game called Dragon Realm, which is probably one of the best outputs so far I've seen from a local model when it comes to building something like this. You create that landing page that shows you a second ago. It's actually pretty good for coding out local mini apps as well. And you can see it coding right here. So it's quite fast when we use it directly. We can also control it with our voice using the Agent Factory and everything that we build gets saved to our workspace too. Now, if you want to learn more about what Agent A1 is, so it's a 35 billion parameter mixture of Agent Sugentic model.
It came from Intern Science, which is a lab in Shanghai.
There's a few different versions of the model. This just dropped 24 hours ago, and it's designed for agentic tasks. That means, for example, you could plug it locally into something like Hermes as well and use it locally as a coding agent. If you're wondering who are Intern Science, so they describe themselves as the open-source hub of AI for Science Center at Shanghai AI Laboratory. They've created quite a few models, so Intern Agent as well as something else they've built. And this is designed for long horizon trajectories as well. So it can work on long horizon tasks according to this.
Now, how does it perform on the benchmarks? I always think like test yourself, particularly when it comes to local models. But according to these benchmarks, for example, if we look at HLE, it's outperforming Quen 3.6 and Step 3.5 Flash, Kimi, DeepSeek and ChatShift are not too far off on that.
And it's particularly good at science benchmarks too. So actually, if you look at the A1 projects page, they say it's optimized for long horizon search, engineering, scientific research, instruction following and tool cooling. Now, if we go back to that task we just gave it, we've got the keyboard here, we can switch the volume, we can change the oscillator.
If you want to see some other stuff that will be built with it, the landing page was pretty nice. It created like this basic sort of neon snake game. Bear in mind, it's not a frontier model, so it's not like going to compete with Fable 5 or something crazy like that, but it can build basic mini stuff. So for example, like mini apps, landing pages, small games, etc. Also created this quite nice solar system task here. So this is an example of a 3D visualization from a solar system. It built that out, no problem, which is pretty amazing itself. It's also 256K context window, which is enough for most tasks. You're not going to hit the token context window for most tasks, and it's designed for agentic reasoning, tool use, long context, and instruction following. So it's a 35 billion parameter model, only activates three billion at time because it's a mixture of experts, and also in terms of speed, it actually ran faster than Gemma 4 on MOX when we tested out. So it ran at 95 tokens per second, fully local on a Mac Studio M4, compared to Qwable which runs way slower, and Gemma 4 Coder way, way slower. In terms of how it performs on the benchmarks versus everything else that we tested, scored 4.8 out of 10 versus Qwable at 7.14. So this is still the strongest model I've tested, but it is a lot slower. And then we've got Gemma 4, which scored lower than Agent A1. But you can also check out goldiebench.com if you're not sure, if you want to check it out yourself and have a look at the builds we created. Now, if you want to run it locally, you can just run it with a couple of commands here. Now, you also might be wondering, okay, what are the strengths and what are the weaknesses of it? So it's tuning is really fine tuned at search tools and science. So it's not really going to be good for like visual stuff, like UI and that sort of thing. This is more for research and agenda tasks. In terms of the strengths, it's agent tuned. So it claims state of the art on seal, zero, long horizon search, IF bench instruction following and browse comp.