China’s NEW Meituan LongCat 2.0 Tested!

**Julian Goldie** (0:00)
Today, we have a brand new update from a Chinese model that's open source. It is called LongCat 2 is here. And this is actually the full model behind AoAlpha. So if you're familiar of AoAlpha, which was a free API, you can actually use it with Hermes. You could plug it into Cloud Code before. It was not bad at all. It is an agentic model. And this was actually revealed as LongCat 2.0. So this is now officially been released and you can get access to it. And you can see the full details right here. So it's got LongCat, Sparse Attention, Zero Compute Experts, MOPD.
Stacks up not badly on the benchmarks here. If you're wondering how to compare against everything. So Terminal Bench 2.1, it holds its own with these other models. As you can see right here, this is compared against Opus 4.6, 4.7 and 4.8. Now, obviously, Opus 4.8 is crushing it by a long way, as you can see here. The other thing that I noticed is if you're using this directly on the website. So if you go to LongCat and then you go to the API section, it looks like you can't use the API unless you have some sort of Chinese setup. So, for example, here, if we try and get a pack. It's broken anyway, look at that.
But if you want to use it, I've been testing out. So I'll show you what we've done so far with it and how it works and how it performs here. So let's have a look at some of the stuff we built with this.
How it performs and benchmarks. Some of it was good, some of it was not so good. So this is a game we created called Dragon Realm. It's not bad. I mean, the graphics are not bad. This is nowhere near the same standard as Glm 5.2, Opus 4.8 or anything like that. But I don't think it's designed to be. I think it's just designed to be a cool model that you can test out. This one was not bad at all as well. So it's kind of like a Skyrim style open world game. As you can see, it runs pretty smoothly. The graphics are not great. And then here's another one. This is probably the best output that I saw so far. And we can compare it on Goldie Bench against a bunch of other examples in a second. But again, it's still pretty basic and buggy, right? Like what is going on here? It just goes completely black. And then we have another example right here. So overall, on my tests, I wouldn't say it's that impressive.
We ran it through Goldie Bench. I'll show you how it compares versus our models in second.
It is a 1.6 open source, 1.6 trillion parameter model.
And the other thing to note here, and this is probably the biggest update about it.
As you can see from this tweet by Robin, is that number one, it's open source, but number two, it was actually built on a different chip. So this was built on China's Meituan.
And this was trained without a single Nvidia chip. Now, if you're wondering who are Meituan, they're basically like China's version of DoorDash. This is pretty crazy. I mean, we saw this with Xiaomi as well earlier this year, where like, you know, everyone's getting involved in AI. Everyone's bringing out their own models. Companies that we don't expect to bring out their own models come out. But yeah, it's pretty cool. I mean, like, fair play to them for bringing out a model like this. Would I say it's the best model I've ever used? Probably not. But it's fun to play with, fun to test out, et cetera. The way that I actually used it, because I couldn't get access to the API and see like it's it's not even available to top up yet, is you can actually go to the chat section and just start using it there. They've also got a full breakdown of how it works step by step, as you can see on the website. Also on the evaluations, this is pretty interesting. GPT 5.5 is only slightly above on Terminal Bench versus LongCat and then on SWE-Bench Pro, LongCat is actually outperforming GPT 5.5. Now again, I've shown you my tests. Do I personally think that it's better than GPT 5.5 on my own benchmarks? Probably not. And Opus 4.8 is beating them all, but you would expect that anyway. Now, I think probably the most relevant side by side comparison is comparing it against GLM 5.2, because that is another open source project that just came out of China recently and is very cheap on the coding plan. And also it's designed to be agentic as well. So, for example, we have a look at this Crypt game. This is the one created from GLM 5.2. It's pretty dark and it is a little bit buggy, but it works. And it's actually, you know, interesting and quite useful. If we compare that versus the option from LongCat here, you can see that this is very limited. It has nothing going on in the game and it's super buggy, right? You can just walk through the wall.