Meituan Unveils LongCat-2.0: A 1.6T-Parameter Model for Agentic Coding
Meituan has introduced LongCat-2.0, a groundbreaking large-scale Mixture-of-Experts (MoE) language model designed for agentic coding tasks. With 1.6 trillion parameters and a native 1-million-token context window, the model aims to revolutionize code understanding, generation, and execution within agent systems. Developed entirely on domestic AI ASIC superpods, LongCat-2.0 promises efficiency and stability, marking a significant leap from its 560B predecessor, LongCat-Flash.
Architecture Breakthroughs
The model’s design prioritizes cost-effective scalability through four key innovations. Zero-computation experts streamline processing by routing simple tokens like punctuation to minimal compute paths, while complex tokens engage higher-capacity experts. A PID controller dynamically adjusts expert activation, enabling a flexible 33B–56B activation window. The ScMoE architecture further boosts throughput, and LongCat Sparse Attention (LSA) reduces memory demands by selecting relevant tokens, achieving near-linear scaling for the 1M-context window.
Performance Benchmarks
Meituan claims LongCat-2.0 excels in software engineering tasks, outperforming GPT-5.5 on SWE-bench Pro (59.5 vs. 58.6) and matching Gemini 3.1 Pro’s capabilities. However, it trails leading systems in broader benchmarks like FORTE and BrowseComp. The model also integrates a 135B N-gram embedding module to capture local token relationships, reducing memory I/O during large-scale decoding.
Domestic Hardware & Stability
Training and serving ran exclusively on China’s AI ASIC superpods, a move that highlights the model’s compatibility with non-Nvidia infrastructure. Meituan emphasizes stability, citing no rollbacks or performance spikes during pretraining—a critical advantage in less mature tooling environments. For deployment, the model employs 6D parallelism and prefill-decode architectures to minimize latency.
LongCat-2.0 underscores Meituan’s push into advanced AI-driven workflows, blending technical innovation with strategic hardware reliance. While its performance in niche coding tasks is promising, broader industry benchmarks remain to be validated.
Source: MarkTechPost. AI-assisted editorial synthesis — TechnoExpress.