The model that was topping OpenRouter under a fake name
For weeks, an unbranded model called "Owl Alpha" was quietly climbing OpenRouter's usage charts, eventually pulling in around 559 billion tokens a day and a spot in the platform's global top three. Then Meituan revealed it: that was LongCat-2.0, a 1.6 trillion parameter model the food-delivery giant had trained entirely on domestic Chinese chips, without touching a single Nvidia GPU. It's now released under the MIT license and built specifically for agentic coding.
What makes it different
LongCat-2.0 is a Mixture-of-Experts model, so it never actually runs all 1.6 trillion parameters at once. Depending on the task, it activates somewhere between 33 and 56 billion parameters per token, which keeps inference far cheaper than a dense model of the same size would be. It ships with a native 1 million token context window and something Meituan calls LongCat Sparse Attention, aimed at keeping long codebases and file trees usable without the usual slowdown.
On benchmarks, it's genuinely competitive rather than just big for the sake of being big. It scored 59.5 on SWE-bench Pro versus GPT-5.5's 58.6, and hit 70.8 on Terminal-Bench. In practical agent harnesses, it took the top ranking on Hermes Agent workspace and placed second on Claude Code deployments.
How to actually try it for free
Here's the catch worth knowing upfront: despite the MIT license announcement, full model weights are still listed as "coming soon" on Hugging Face and GitHub. You can't self-host it yet. What you can do today:
- Go to Meituan's own LongCat API platform at longcat.chat and create an account.
- Grab an API key. The platform currently gives 5,000,000 free tokens per day for the LongCat-2.0-Preview model, no payment required to start.
- Point an OpenAI- or Anthropic-compatible client at the LongCat endpoint, or use an agent harness like Claude Code or OpenClaw that already has LongCat integration.
- Alternatively, reach it through OpenRouter, where standard pricing runs $0.75 per million input tokens and $2.95 per million output, with a launch promotion cutting that to roughly $0.30/$1.20 for a limited window.
If you'd rather skip the API setup entirely, open-source agent tools like [Kilo](https://questloops.com/tools/kilo) already let you plug in alternative model backends, including OpenAI-compatible endpoints like LongCat's, without writing your own integration.
LongCat-2.0 vs the free tier you're used to
| LongCat-2.0 (Meituan API) | Typical free-tier coding model | |
|---|---|---|
| Free daily allowance | 5M tokens/day (Preview model) | Usually message-count limited, not token-based |
| Context window | 1M tokens native | Often 32K-200K |
| Local self-hosting | Not yet (weights pending) | Varies |
| License | MIT (once weights ship) | Varies, often closed |
Worth trying, with a real caveat
If you're already paying for coding-agent tokens, LongCat-2.0 is worth pointing at a side project purely on the benchmark numbers and the free daily allowance from Meituan's own platform. The honest limitation is that "open source" here is currently a promise, not a download. Until the weights actually land on Hugging Face, you're relying on Meituan's hosted API, which means you don't get the self-hosting, fine-tuning, or full data-privacy benefits that usually come with an MIT license. Treat it as a very capable free API today, and an open-weights model tomorrow, not the same thing yet.
