AI Model Report

Open Source · JULY 5, 2026

Meituan ships LongCat-2.0: 1.6T MoE, 1M context, trained end-to-end on Chinese ASICs

Meituan open-sourced LongCat-2.0 on Hugging Face and GitHub under MIT, unmasking the 1.6-trillion-parameter MoE that had been leading OpenRouter as 'Owl Alpha' — and the first trillion-parameter system pretrained and served entirely on a 50,000-card domestic ASIC cluster.

By Lars Iverson · Open source & model weights · July 5, 2026

Meituan released LongCat-2.0 on Tuesday under an MIT license, dropping weights simultaneously on Hugging Face, GitHub, and its own longcat.ai portal, and in the process confirming what OpenRouter's leaderboard-watchers had suspected for weeks: the anonymous "Owl Alpha" model that had been sitting atop the global developer charts for roughly two months was a Chinese food-delivery company's 1.6-trillion-parameter mixture-of-experts, trained end-to-end on domestic silicon.

The headline architecture: 1.6T total parameters with a nominal ~48B activated per token, ranging dynamically between 33B and 56B depending on query complexity via a Zero-Computation Experts gating framework. Native context is 1 million tokens, threaded through a linear-complexity attention variant Meituan calls LongCat Sparse Attention with Streaming-aware Indexing. A further 135B N-gram Embedding parameters are carried over from the predecessor LongCat-Flash-Lite.

The compute story is the one that'll echo. Reuters and SCMP both report LongCat-2.0 as the first trillion-parameter system to complete pretraining and inference on a 50,000-card cluster of Chinese-made AI ASIC superpods, with no Nvidia hardware at any stage. DeepSeek V4-pro, at comparable scale, still relied on domestic silicon only for inference. The vendor behind Meituan's ASICs remains unnamed in the official disclosures, which is itself a data point about how sensitive the supplier relationships have become.

The training run consumed "millions of accelerator-days across more than 35 trillion tokens, with no rollbacks or irrecoverable loss spikes," per the Hugging Face model card. That claim, if it holds, is the more analytically interesting one. Cluster-level training stability on non-Nvidia silicon at trillion-parameter scale is the specific bottleneck that Washington's 2022 and 2023 export-control regimes were designed to preserve as a US advantage.

Benchmarks land where the OpenRouter usage data already implied they would. SWE-bench Pro: 59.5, ahead of GPT-5.5 at 58.6. Terminal-Bench: 70.8. SWE-bench Multilingual: 77.3. As Owl Alpha, the model was handling roughly 559 billion tokens a day and about 10.1 trillion monthly on OpenRouter, a 242% month-over-month jump, taking first on Hermes Agent workspace, second on Claude Code deployments, and third across OpenClaw environments.

Reference serving, per the model card, runs on SGLang with 16× H20 accelerators against the meituan-longcat/LongCat-2.0-FP8 checkpoint. Developers pulling weights this week will note the small print on the GitHub and Hugging Face repos: "Model weights coming soon — stay tuned!" reads a stale notice VentureBeat captured at publication, though the FP8 checkpoint is live.

The strategic read is straightforward. The 2019 Huawei entity-listing and the successive chip-export tightening under two US administrations were premised on a compute chokepoint that would slow Chinese frontier training by years. LongCat-2.0 is the first open-weights artifact that says, on the record and reproducibly, that the chokepoint has a workaround at trillion-parameter scale, shipped by a company whose day job is coordinating restaurant couriers.

Sources