Meituan's LongCat-2.0 Secretly Topped OpenRouter for Two Months. Now It's Open Source

For two months, a mystery model called “Owl Alpha” sat near the top of OpenRouter’s leaderboards, and nobody outside Meituan knew who built it. On June 30, 2026, Meituan came forward: Owl Alpha was LongCat-2.0, a 1.6-trillion-parameter coding model, and the company open-sourced the whole thing under an MIT license the same day. The bigger claim buried in the announcement is the one that matters more for the AI industry’s supply chain: Meituan says it trained and serves this model entirely on Chinese-made chips, with no Nvidia or AMD hardware anywhere in the pipeline.

The stealth launch

Before anyone knew what Owl Alpha was, it was already competing. On OpenRouter, a service that routes requests across dozens of AI providers so developers can compare and switch between them, the anonymous model had climbed to first place on the Hermes Agent workspace leaderboard and second place on the Claude Code leaderboard. That’s an unusual way to launch a model: most labs announce first and let the benchmarks follow. Meituan let the benchmarks speak before attaching a name to them, which is either confidence or a very effective marketing trick, and probably both.

What’s actually under the hood

LongCat-2.0 is a Mixture-of-Experts model: 1.6 trillion total parameters, but only an average of about 48 billion are active for any given token, which is what makes a model this large computationally feasible to run at all. It ships with a native 1-million-token context window, built on what Meituan calls LongCat Sparse Attention, aimed at reasoning over large codebases without the context getting truncated or summarized away.

On the benchmarks that matter for a coding model:

LongCat-2.0 lands at 59.5% on SWE-bench Pro, ahead of GPT-5.5 (58.6%) and Gemini 3.1 Pro (54.2%), but still behind Claude Opus 4.7 (64.3%) and Opus 4.8 (69.2%). It also posts 70.8 on Terminal-Bench 2.1 and 77.3 on SWE-bench Multilingual. That’s not a model that tops every leaderboard. It’s a model that’s genuinely competitive with most of the field while trailing only the current frontier leader, which is a very different story than the open-source-model-as-afterthought narrative from a couple of years ago.

The part that matters beyond the benchmark table

The benchmark numbers are good, but they’re not the most consequential detail in the announcement. Meituan says LongCat-2.0 was trained on a roughly 50,000-card domestic compute cluster, using chips manufactured in China, with zero reliance on Nvidia A100s, H100s, or AMD’s MI300-class accelerators anywhere in training or inference. If that holds up to scrutiny, it’s the first reported case of a trillion-parameter-scale model built and served end-to-end without the specific hardware that export controls have spent years trying to keep out of reach.

That matters beyond the geopolitics angle. It’s evidence that the ceiling on frontier-scale AI training isn’t as hard-gated by a single hardware supply chain as the last few years of export-control policy assumed. Whether that’s read as a policy failure or an inevitability depends on who you ask, but either way, it changes the planning assumptions for anyone betting on chip scarcity as a durable constraint on who can build frontier models.

What this means if you’re choosing a coding model

Most teams evaluating AI coding tools right now are picking between a small number of proprietary APIs, weighing benchmark scores against pricing and data-handling terms. LongCat-2.0 adds a genuinely different option to that list: an MIT-licensed, self-hostable model that’s competitive with, though not ahead of, the frontier proprietary systems.

That matters most for two kinds of teams. First, anyone with data residency or compliance constraints that make sending proprietary code to a third-party API a hard no, rather than a preference, now has a model worth evaluating that they can run on infrastructure they control. Second, anyone doing high-volume agentic coding work, where API costs scale with usage, has a credible self-hosted alternative to weigh against per-token API pricing, assuming they have the infrastructure to serve a trillion-parameter model at reasonable latency, which is itself a nontrivial undertaking.

It doesn’t replace the case for frontier proprietary models if your priority is the single best score on every benchmark. Claude Opus 4.8 still leads SWE-bench Pro by a wide margin. But “wide margin behind the leader” and “not a viable option” are different claims, and LongCat-2.0’s numbers put it firmly in the first category, not the second.

If you’re mapping out which coding tools and models actually fit your team’s workflow rather than just chasing the top benchmark score, our guide to the best AI coding tools walks through that evaluation from the tooling side, and pairs well with treating a model choice like LongCat-2.0 as one more input to that decision rather than a headline to react to.

The takeaway

LongCat-2.0 isn’t the best coding model available today, and Meituan isn’t claiming it is. What it demonstrates is that a credible, open-weight alternative to the usual handful of frontier vendors now exists, built on infrastructure nobody expected to be capable of it two years ago. For teams with real constraints on where their code and API calls can go, that’s a more useful development than another leaderboard headline.