If you've used any AI coding tool in the last two months — Cursor, GitHub Copilot, something that describes itself as "AI-powered" — there's a real chance some of your code suggestions came from an AI nobody knew was Chinese.

Meituan confirmed on June 30 that "Owl Alpha" — the anonymous model that had been ranking first on Hermes Agent, second on Claude Code integrations, and third on OpenClaw by call volume on OpenRouter — is LongCat-2.0. 1.6 trillion total parameters. MIT license. Trained entirely on Chinese domestic chips, with no NVIDIA H100s, and nothing that falls under US export controls.

The actual downloadable weights still say "coming soon" on Hugging Face.

Why the two-month anonymous run matters

OpenRouter is an API aggregator — think of it like a taxi dispatch system for AI models. When your coding tool sends a request and doesn't specify which model to use, OpenRouter picks whichever model scores best or cheapest for that task, automatically. For two months, the one it kept selecting was something called Owl Alpha.

The builders whose tools were calling Owl Alpha during that window just knew: the coding tasks work, the completions are good, keep going. Nobody in the routing layer was steering traffic toward a Chinese model. The algorithm sent requests based on performance alone.

Owl Alpha ranked first on the Hermes Agent framework — the framework that tests multi-step agentic coding tasks, the ones where a model has to plan what it's doing, execute a sequence of steps, and not fall apart in the middle. That's earned ranking. OpenRouter's algorithm doesn't have a country of origin preference.

What 1.6 trillion parameters on Chinese chips actually proves

Mixture-of-Experts architecture (MoE for short — the AI splits work across specialized mini-models depending on the task) means the model doesn't use all 1.6 trillion parameters every time you prompt it. It activates about 48 billion per token. The 1.6 trillion is the total capacity that gets routed to specialists depending on what you're asking. That's the same design DeepSeek used, and it's what makes the model practical to run without a cluster of the most expensive NVIDIA hardware.

The chip independence claim is load-bearing. US export controls since late 2022 have restricted NVIDIA's most capable AI training chips from reaching China. The theory was clear: without H100s and their successors, China couldn't train frontier-grade models. LongCat-2.0 is the largest direct challenge to that theory yet. DeepSeek V3 and V4 made the argument at smaller scale. Meituan is making it at 1.6 trillion parameters, trained end-to-end on domestic hardware.

This doesn't mean the export controls failed completely. They may have added years, or billions in cost, or forced architectural choices that shaped the result. But "no H100 = no frontier AI" is no longer holding at this scale.

¹ LongCat-2.0's Terminal-Bench version is not specified. Sonnet 5 and Opus 4.8 scores are from Terminal-Bench 2.1. Direct comparison is approximate.

Source spread

• Meituan / LongCat AI — official model page — [builder]. Primary source for specs, architecture, and API access. Benchmark numbers are self-reported without independent methodology documentation as of this writing.
• VentureBeat — Meituan open-sources LongCat-2.0 — [builder]. Best on OpenRouter context and what chip independence means commercially.
• Decrypt — LongCat-2.0: the stealth AI topping OpenRouter all along — [builder]. Best on the Owl Alpha backstory and what two months of anonymous top-ranking traffic actually signals.
• Lifeboat News — first frontier LLM on Chinese domestic chips — [skeptic]. Strongest on the policy significance and what this means for the US export control argument.

Pros & cons

What's real:

The OpenRouter run is the most credible data point in this release. Two months of real developer traffic, a routing algorithm with no preference for Chinese models, and a consistent top-three finish across major agent frameworks. Owl Alpha earned that ranking by performing on actual tasks — not on a synthetic benchmark Meituan designed.

The chip-independence demonstration is real, and it's the largest to date. Training 1.6 trillion parameters on domestic hardware, without restricted NVIDIA chips, puts a concrete marker in the ground. DeepSeek started this argument. Meituan is extending it.

MIT license is genuinely permissive. Commercial use, modification, redistribution, no restrictions. More open than Meta's Llama license agreements, which have commercial-use carve-outs for large organizations.

1M token context window for a coding model is practical for real codebases. Most production repositories overflow smaller windows. LongCat-2.0 was designed specifically for the workloads where that matters.

What deserves a side-eye:

The SWE-bench Pro gap is real and slightly confusing. 59.5 is below Sonnet 5's 63.2 and substantially below Opus 4.8's 69.2 on the same benchmark. If the model was genuinely topping OpenRouter's Hermes Agent rankings over two months, you'd expect the canonical coding benchmark to be closer. The OpenRouter traffic story and the SWE-bench Pro number tell slightly different stories about where this model actually sits.

The weights problem. "Open source" in 2026 has split into two meanings: "weights released" (you can download and run it yourself) and "license declared open" (they say it's open but the files aren't there yet). LongCat-2.0 is currently in the second category. The model is available through Meituan's API and through OpenRouter, but the Hugging Face and GitHub pages still say "model weights coming soon" with no date given.

Benchmark version mismatch: Meituan reports a Terminal-Bench score without specifying which version. The Sonnet 5 and Opus 4.8 numbers I have are from Terminal-Bench 2.1. These may or may not be the same benchmark run. I've flagged that in the table footnote; treat the comparison as directional.

What to do about it

For everyday AI users and builders both:

• You may already have run LongCat-2.0 if you use Cursor, Claude Code, or any OpenRouter-integrated tool. Check your tool's model logs if you're curious which model handled specific completions over the last two months.
• Try the API directly if you want to evaluate it now. longcat.ai has OpenAI-compatible endpoints — the same format as calling the OpenAI API — so it drops in to most existing setups with a one-line change.
• Don't call it "open source" yet. The MIT license is real. The downloadable weights aren't. If your use case requires self-hosting — for privacy, cost, or compliance reasons — wait for the weights before planning around it.
• If the weights ship under MIT, this becomes the most permissive frontier-grade coding model available for commercial self-hosting. Worth keeping in the evaluation queue. The weights landing date is the event to watch.

Further reading

• Meituan LongCat AI — official model page — specs, API docs, benchmark claims
• Meituan LongCat on Hugging Face — where the weights will land when they ship
• VentureBeat — Meituan open-sources LongCat-2.0, leading OpenRouter
• Decrypt — LongCat-2.0: the stealth AI topping OpenRouter all along
• Lifeboat News — Meituan trains first frontier LLM on Chinese domestic chips