Meituan, the Chinese technology company best known for food delivery and local commerce, released LongCat-2.0 on June 30 — a 1.6-trillion-parameter Mixture-of-Experts model available under an MIT License.
Built Without Western Chips
The model was trained on more than 50,000 domestically produced Chinese ASICs (application-specific integrated circuits), without relying on Nvidia GPUs or other Western semiconductor hardware. It marks a significant milestone in China’s effort to develop frontier-class AI capability independent of Western chip supply chains. Training consumed more than 35 trillion tokens.
Near-Frontier Benchmark Performance
LongCat-2.0 scored 59.5 on SWE-bench Pro, a benchmark for real-world software engineering tasks, edging ahead of GPT-5.5’s 58.6. Additional results: 70.8 on Terminal-Bench 2.1, 77.3 on SWE-bench Multilingual, and 73.2 on FORTE.
The model natively supports a one-million-token context window, enabling it to process entire large codebases or lengthy documents without truncation. Architecturally, it uses a LongCat Sparse Attention mechanism, an N-gram Embedding layer, and a Zero-Compute Experts framework for improved efficiency.
From “Owl Alpha” to Open Source
Before the public announcement, LongCat-2.0 had been quietly deployed on OpenRouter under the anonymous name “Owl Alpha,” where it ranked among the top three most-used models globally by developer traffic. Developers adopted it at scale before Meituan disclosed its identity — providing the company real-world usage data ahead of the official launch.
Commercially Permissive License
The MIT License permits the model to be incorporated into proprietary software and redistributed without requiring derivative works to be open-sourced — one of the least restrictive licenses applied to a model of this scale. Model weights are available on Hugging Face and ModelScope; API access is available through Meituan’s platform at longcat.ai.