Woofun AI reports that Meituan has open-sourced LongCat-2.0, a 1.6 trillion parameter Mixture of Experts model supporting 1 million token context. The model completed pre-training on 35 trillion tokens using a cluster of over 50,000 domestic AI chips, marking the first trillion-parameter model to undergo full training and inference on domestic hardware.

The update introduces LongCat Sparse Attention to optimize memory overhead via flow-aware and hierarchical indexes. It also integrates a 135 billion parameter 5-gram embedding module to enhance local context representation. On benchmarks like SWE-bench Pro, LongCat-2.0 performance rivals or exceeds several mainstream closed-source models.