01What is Meituan LongCat 2.0?

Meituan LongCat 2.0 is a new open-source AI model from China's Meituan company. It was released after being the model behind a free alpha API called "Hermes" that you could plug into code. In this episode, Julian Goldie tests it and shows how it works.

The model is huge, with 1.6 trillion parameters. It uses a special "LongCat sparse attention" and "zero compute experts" mechanism. It is designed to be a general model that you can use for chat, coding, and creating games.

The biggest news is that it was trained without a single NVIDIA chip, using China's Meichuan chip instead.

Meichuan is like China's version of DoorDash, a delivery company. This shows that unexpected companies are now entering the AI space. The model is open source, meaning anyone can access the code and weights.

Julian notes that on benchmarks, it holds its own against other models, but in his own tests, it is not the best.

02Main Features of LongCat 2.0

  • Open source: Anyone can get access to the model and its details.
  • 1.6 trillion parameters: This makes it a very large and capable model.
  • LongCat sparse attention: A technical innovation that helps the model process information efficiently.
  • Zero compute experts: Another technical feature that reduces computing needs.
  • Built on Meichuan chip: Trained without any NVIDIA chips, which is a big deal for chip independence.
  • Agentic design: It is designed to work as an agent, similar to other modern models.

The model also has a chat interface that you can use directly on the website if the API is not accessible. The website includes a full step-by-step breakdown of how the model works. Evaluations show that on Terminal Bench 2.1, GPT 5.5 is only slightly above LongCat 2.0.

On SWE Bench Pro, LongCat 2.0 actually outperforms GPT 5.5.

However, Julian points out that on his own GoldieBench tests, he personally does not think it is better than GPT 5.5. Opus 4.8 beats all of them, but that is expected. The model is not designed to be the best, but rather a cool model to test.

03How to Set Up and Get Started with LongCat 2.0

  1. Go to the LongCat website and navigate to the API section.
  2. Try to get a pack or top up your account. Note that it looks broken if you are not in China.
  3. If the API is not accessible, go to the chat section on the website and start using it there directly.
  4. You can also read the full breakdown on the website to understand how it works step by step.
  5. For the API, you may need a Chinese setup or payment method.

Julian could not use the API because it was not available to top up yet. So he used the chat interface for his tests. The model is open source, so you could also run it locally if you have the hardware.

To compare it with other models, you can use it in the same way you would use GLM 5.2 or any other open-source model. The setup is straightforward if you stick to the chat interface. For developers, the open-source nature means you can download the weights and integrate it into your own projects.

04How Julian Tested LongCat 2.0 in the Episode

Julian used LongCat 2.0 to create simple games to test its capabilities. He showed several examples: Dragon Realm, a Skyrim-style open world game, and a Crypt game. The games vary in quality and smoothness.

For Dragon Realm, the graphics are not bad, but it's nowhere near the standard of GLM 5.2 or Opus 4.8. The Skyrim-style game runs pretty smoothly, but the graphics are not great. One output he called the best went completely black, showing bugs.

Another game had issues like walking through walls.

He also compared LongCat 2.0 to GLM 5.2. On most game benchmarks, GLM 5.2 won clearly. For example, the GLM 5.2 version of Dragon Realm was more playable and had better graphics.

The Voxel Craft game from GLM 5.2 worked perfectly, while the LongCat version was not at the same level.

Overall, Julian says the model is fun to test but not the best. He recommends GLM 5.2 as the strongest open-source model he has tested.

05Benchmark Results and Comparisons in the Episode

In the episode, Julian shows benchmark results from the LongCat 2.0 website. On Terminal Bench 2.1, LongCat 2.0 holds its own against models like Opus 4.6, 4.7, and 4.8. Opus 4.8 is crushing it by a long way, but LongCat is competitive with the others.

On SWE Bench Pro, LongCat 2.0 actually outperforms GPT 5.5. This is a notable result, though Julian notes that his own GoldieBench tests suggest otherwise. He says, "Do I personally think that it's better than GPT 5.5 on my own benchmarks?

Probably not."

Julian also ran his own GoldieBench tests, comparing game creation abilities. He found LongCat 2.0 to be "basic and buggy" in many cases. The GLM 5.2 versions of the games were consistently better and more interesting.

In the end, Julian says he will not be switching to LongCat 2.0 anytime soon, but he wanted to test it and show what it looks like.

06Tips, Limits, and Mistakes to Avoid with LongCat 2.0

  • API access is limited: If you are not in China, you may not be able to use the API. Stick to the chat interface.
  • Expect bugs: The model is still basic and buggy, especially in game creation. Be prepared for issues like black screens or walls you can walk through.
  • Manage expectations: It is not designed to be the best model. It's more of a fun, experimental model.
  • Check the website: The full breakdown of how it works is available on the website, so read it to understand its capabilities.
  • Compare with GLM 5.2: If you are choosing between open-source models, GLM 5.2 is stronger in Julian's tests.

Julian also advises that the model may not be suitable for serious production use yet. But it's worth testing out for fun or research. The fact that it was trained without NVIDIA chips is a significant technical achievement, even if the performance isn't top-tier.

07Who Should Use LongCat 2.0 and Why It Matters

LongCat 2.0 is for developers and AI enthusiasts who want to experiment with new open-source models. It is especially interesting because it shows how Chinese companies, even delivery companies like Meituan, are getting into AI. This is important because it diversifies the AI landscape beyond the usual players like NVIDIA and OpenAI.

It also matters for chip independence. Training a 1.6 trillion parameter model without a single NVIDIA chip is a major milestone. It shows that alternatives to NVIDIA are possible.

While Julian wouldn't use it for his own projects, he acknowledges it's fun to play with. If you're in the AI Profit Boardroom community, you could plug it into the mission control dashboard like they did with GLM 5.2, though he prefers GLM 5.2.

Overall, LongCat 2.0 is a sign of the times. Every company seems to be bringing out its own model, and this one is worth a look for its technical innovations, even if the performance isn't the best.