There is a certain elegance to how Meituan validated LongCat-2.0. Rather than arriving with a press release and a benchmark table, the model spent two months on OpenRouter under the name “Owl Alpha” — anonymous, unannounced, and accumulating usage. By the time Meituan stepped forward on June 30, Owl Alpha had already earned the top slot on Hermes Agent by monthly call volume, second place on Claude Code, and third across OpenClaw deployments — accumulating around 10 trillion monthly tokens in aggregate. It was, in effect, a blind taste test conducted on real developer workloads.
The performance claims that now accompany the official release carry a bit more weight because of that history. Independent users were routing substantial agent traffic through this model without knowing what it was, and it kept earning their calls.
What it is
LongCat-2.0 is a sparse mixture-of-experts architecture with 1.6 trillion total parameters, activating roughly 48 billion per token — swinging between 33B and 56B depending on query complexity. The native context window is 1 million tokens, enabled by LongCat Sparse Attention (LSA), which Meituan describes as an evolution of DeepSeek Sparse Attention. The key engineering problem LSA addresses is that fine-grained sparse attention tends to accumulate quadratic scoring costs and memory fragmentation as context grows; their solution uses streaming-aware indexing to keep that under control.
For n-gram tokenization, Meituan introduced a new embedding system that treats multi-word phrases as single semantic units, so the model handles entities like “New York City” as a concept rather than three independent tokens. This matters most at very long contexts where repeated entity references compound the benefit.
Expert routing is dynamic rather than fixed: a “zero-computation expert” mechanism routes simple token passes through fewer active experts while reserving the full 56B activation budget for difficult queries. This is roughly analogous to how DeepSeek-V3 handles load balancing, but Meituan’s variant ties activation count to estimated query difficulty rather than purely to load.
Training used 6D parallelism: tensor, context, expert, data, pipeline, and embedding parallelism simultaneously, distributed across a cluster of more than 50,000 domestic AI ASICs. The inter-chip communication stack ran on Huawei’s HCCL — the Chinese analogue of NVIDIA’s NCCL — and the training run covered more than 30 trillion tokens, including hundreds of billions at approximately 1M-token context lengths.
On SWE-bench Pro the model scores 59.5, putting it ahead of Gemini 3.1 Pro and GPT-5.5 on coding tasks. Reasoning benchmarks tell a different story: it trails on IFEval, IMO-AnswerBench, and GPQA-Diamond, which puts it firmly in the “strong coder, middling reasoner” quadrant — similar positioning to the original Kimi K2.
The chip independence claim
Training a 1.6T parameter model without a single NVIDIA accelerator is the most strategically significant thing about this release. If accurate, it is evidence that Chinese AI labs can operate at frontier scale on domestic compute — something that export controls since 2022 were meant to foreclose.
But there are real gaps in the disclosure. The chip manufacturer is not named anywhere in the official announcement. Meituan refers only to “domestic AI ASICs” and the Huawei HCCL dependency is the main public signal pointing toward Atlas-950 accelerators — that connection is inferred, not confirmed. Training duration is also not disclosed, which means observers cannot estimate the efficiency delta relative to an NVIDIA cluster of similar size. One analysis estimated the total cost of a 50,000-card domestic cluster at roughly 10 billion yuan over three years including depreciation and power; Meituan has not commented on those figures.
The “open-source” framing deserves scrutiny too. At announcement the model weights were listed as coming soon on Hugging Face, with only inference code available immediately. Meituan’s MIT license claim is real — the license header is there — but a license on code without accessible weights is not the same as the DeepSeek-V3 style release that let anyone download and run a full model. The weights are now listed as available, though the training data composition remains entirely undisclosed.
None of this invalidates the release. The anonymous OpenRouter deployment is the most important external validation available, and the benchmarks appear consistently reported across multiple evaluators. But “trained on domestic chips” is a strategic claim that deserves more transparency than Meituan has offered so far. The engineering achievement and the PR achievement are both present here; they are not the same thing.
What it means
For developers, LongCat-2.0 is worth running on long-context agentic coding tasks. The 1M context is native rather than bolted on, the MoE efficiency means token costs stay reasonable, and the OpenRouter usage history provides some assurance that it holds up under real workloads rather than only curated benchmarks. API pricing at $0.75/$2.95 per million tokens (with a launch promo at $0.30/$1.20) puts it below Sonnet 5 for comparable capability on coding tasks.
For anyone watching the hardware race, this is a meaningful data point regardless of the missing chip details. Chinese labs trained Kimi K2.6 and GLM-5.2 on domestic hardware earlier this year; LongCat-2.0 pushes the demonstrated scale to 1.6T parameters on a single training run. Whether or not the Atlas-950 is as efficient as an H100, the capability to complete a run at this size without imported accelerators matters for the longer trajectory.
The Owl Alpha period — two months of anonymous deployment accumulating top rankings — suggests Meituan ran a deliberate experiment in market validation before making the claim official. That kind of discipline is not common in model releases, and it adds credibility to the performance story even where the hardware story remains opaque.