The Chinese food-delivery giant Meituan has thrown an unexpected punch in the agentic-AI race. The company open-sourced LongCat-2.0, a 1.6-trillion-parameter model built specifically for agentic coding — and revealed that it was the secret engine behind "Owl Alpha," the anonymous stealth model that had quietly climbed to the top of global developer charts over the preceding two months. More striking still: the entire system was trained without a single NVIDIA chip.

The Model Behind the Mystery

For weeks, developers on OpenRouter had been routing enormous volumes of work through an unbranded model called Owl Alpha, unaware of who built it. The usage was staggering — roughly 10.1 trillion tokens per month, averaging 559 billion tokens a day, a 242% month-over-month surge that pushed it into the platform's global top three. When Meituan finally stepped forward, it unmasked Owl Alpha as LongCat-2.0, publishing the model on GitHub and Hugging Face under a permissive MIT license that allows commercial use.

By the time the reveal came, the model had already topped the Hermes Agent workspace, ranked second across Claude Code deployments, and placed third in international agent environments. In other words, developers had voted with their tokens long before they knew whose model they were using — an unusually pure signal of real-world capability rather than launch-day hype.

Purpose-Built for Agents, Not Chat

LongCat-2.0 is not a general chatbot repurposed for code. It is a Mixture-of-Experts system with 1.6 trillion total parameters, activating roughly 48 billion per token — the active count flexing between 33 billion and 56 billion depending on how hard the query is. That sparsity is the central trick: it lets the model carry frontier-scale knowledge while running at the cost of a far smaller dense model.

The design targets the specific demands of agentic software work:

  • Repository-level edits rather than isolated snippets.
  • Automated, multi-step task execution inside agent loops.
  • Long-horizon workflows sustained across a native 1-million-token context window, enabled by a custom mechanism the team calls LongCat Sparse Attention that keeps compute costs linear rather than quadratic.

On Meituan's own benchmarks, LongCat-2.0 scores 59.5 on SWE-bench Pro — edging out GPT-5.5 at 58.6 — and 70.8 on Terminal-Bench 2.1. As always, vendor-reported numbers warrant independent scrutiny, but the OpenRouter usage data lends them unusual credibility. Promotional API pricing starts at $0.30 per million input tokens with free context-cache hits, a fraction of comparable proprietary systems.

Trained Entirely on Chinese Silicon

The most geopolitically loaded detail is the hardware. LongCat-2.0 is described as the first trillion-parameter model to complete full training and inference without NVIDIA GPUs. The entire run happened on a 50,000-card Ascend 910 cluster inside China, using Huawei's HCCL communication library and in-house parallelism, across more than 35 trillion tokens of pretraining.

That matters far beyond one model. US export controls have been designed to throttle China's access to cutting-edge AI accelerators on the assumption that frontier training requires them. A near-frontier, trillion-parameter agent model trained end-to-end on domestic chips challenges that assumption directly — and does so in the open, where the artifact cannot be recalled.

Why It Matters

LongCat-2.0 lands at the intersection of three trends reshaping AI in 2026. First, agentic coding has become the proving ground for model capability, because writing and executing code across many steps is a concrete test of the multi-step reliability agents need everywhere. Second, open weights from Chinese labs keep compressing the gap with proprietary Western leaders — and an MIT license means enterprises can embed the model directly into closed commercial products with no legal friction.

Third, and most consequentially, the hardware independence reframes the strategic picture. If competitive frontier models can be trained on domestic accelerators, the leverage of chip export controls diminishes, and the center of gravity in open AI continues drifting eastward. For engineering teams, the practical upshot is a genuinely capable, self-hostable agent model with a million-token context and permissive licensing — one they can inspect, fine-tune, and run without routing sensitive code through a foreign API.

One caveat deserves flagging: at least one report questioned whether the full weights were immediately downloadable at launch, so teams planning production deployment should verify current availability before committing. But the direction is unmistakable. A stealth model can no longer hide its origins once it dominates the charts — and increasingly, those origins are open, Chinese, and independent of the hardware the West assumed everyone needed.

Sources