Meituan Releases LongCat-2.0: A 1.6T-Parameter Open MoE Model with Native 1M Context and LongCat Sparse Attention

A company released an open large language model with 1.6 trillion total parameters designed specifically for software engineering tasks like code generation and debugging. The model can process 1 million tokens of context at once through a technique called sparse attention that reduces computational cost from quadratic to linear scaling. Training and serving happened entirely on domestic AI chips rather than Nvidia hardware, demonstrating capability on non-standard infrastructure. According to vendor benchmarks, it scores 59.5 on real-world software engineering tasks and 70.8 on shell execution benchmarks, positioning it as competitive with other frontier models in coding-focused work.