Meituan LongCat 2.0 just officially released — an open-source Chinese agentic model, and the full model that was quietly powering the AoAlpha free API. I put it through real tests (benchmarks, live game builds, and a side-by-side with GLM 5.2), so here's the honest verdict on whether it's worth your time.
📺 Watch: LongCat 2.0 tested — benchmarks, games, and the GLM 5.2 comparison
Want every new model tested and wired into one system? That's the Agent OS inside the AI Profit Boardroom. → Join AIPB
What Is Meituan LongCat 2.0?
LongCat 2.0 is an open-source, agentic model from Meituan — yes, the Chinese "DoorDash" company — and it turns out to be the model behind AoAlpha, the free API some of us were already using with Hermes and Claude Code. It ships with LongCat Sparse Attention, Zero Compute Experts, and MIPD, and it's a 1.6-trillion-parameter model. The standout detail: it was trained entirely on China's own Meituan chips, without a single Nvidia chip — a genuinely notable moment in the AI hardware race.
LongCat 2.0 Benchmarks
On paper it holds its own: on Terminal Bench 2.1 it's competitive with Opus 4.6–4.8 (though Opus 4.8 still wins clearly), and on SWE-Bench Pro it actually edges out GPT-5.5. One access gotcha worth flagging — the official API section looked broken when I tried it, and topping up seemed to require a Chinese setup, so I tested it through the website chat instead.
I Tested It: Building Real Games
Benchmarks are one thing; building with it is another. I had LongCat 2.0 build a few game demos — Dragon Realm, a Skyrim-style open world, and VoxelCraft. The verdict: playable, but buggy and rough around the edges — one build even goes completely black at points. It's clearly not designed to compete at the very top; it feels more like a fun model to experiment with than a serious daily driver.
LongCat 2.0 vs GLM 5.2: The Honest Verdict
This is the comparison that matters, since both are open-source Chinese models. I ran the same game builds on each: GLM 5.2's version of the crypt game, Dragon Realm, the Skyrim-style world, and VoxelCraft all came out cleaner, smoother, and more polished than LongCat's equivalents. LongCat's VoxelCraft build especially wasn't close. Across every side-by-side, GLM 5.2 came out on top.
Should You Use Meituan LongCat 2.0?
My honest take: it's a genuinely cool, free, open-source release, and credit to Meituan for shipping a serious model from outside the usual AI labs. But if you're choosing an open-source Chinese model to actually build with today, I'd still go with GLM 5.2 — it's the strongest open-source option I've tested. LongCat 2.0 is worth playing with, not switching to.
Get Every New Model Tested — And Wired Into One System
I test models like LongCat 2.0 the day they drop and plug the winners straight into the Agent Operating System inside the AI Profit Boardroom — the same dashboard where GLM 5.2 is already running through Claude Code and Hermes. Whatever wins next, you swap it in without rebuilding anything. → Join AIPB.
Frequently Asked Questions
What is Meituan LongCat 2.0?
An open-source, agentic AI model from Meituan, the Chinese delivery company — the model behind the AoAlpha free API, trained without Nvidia chips.
Is LongCat 2.0 better than GLM 5.2?
In my hands-on tests, no — GLM 5.2 outperformed LongCat 2.0 across every game-build comparison.
Is Meituan LongCat 2.0 free?
Yes, it's open-source and free to use, though API access can be tricky outside a Chinese setup — the website chat works as an alternative.
The Bottom Line
Meituan LongCat 2.0 is a genuinely interesting, free, open-source model — trained without Nvidia, decent on some benchmarks, but buggy in real use and outperformed by GLM 5.2 in every head-to-head I ran. Worth testing, not worth switching your daily driver to.
