Meituan LongCat 2.0: A 1.6T Parameter Mixture-of-Experts Model

Model Architecture and Scale

Meituan has introduced LongCat 2.0, a Mixture-of-Experts (MoE) model featuring 1.6 trillion total parameters with approximately 48 billion parameters activated per token. The architecture incorporates a proprietary "LongCat Sparse Attention" mechanism designed to improve efficiency for long-context tasks. Additionally, the model utilizes an n-gram embedding module that accounts for 135 billion parameters, intended to enhance parameter utilization. The developers report that the model was trained on over 35 trillion tokens using AI ASIC superpods rather than standard Nvidia GPU clusters.

Performance and Testing

While the model is marketed for agentic coding, reasoning, and long-context tasks, initial one-shot testing via the public web interface yielded poor results. On the BigCodeBench benchmark, the model scored 21.6%, placing it significantly behind models like DeepSeek V4 Pro and Gemini 3.5 Flash. It struggled specifically with creative coding tasks, often failing to produce functional outputs for simulations and 3D modeling requests. The author notes that these results may be misleading, as the model is likely optimized for multi-step agentic workflows where it can iterate, edit files, and run commands, rather than the single-turn responses tested on the web platform. Access to the model's API and specialized coding tools remains restricted to users in China, and the full model weights have not yet been released on Hugging Face.