Meituan released LongCat-2.5 on Monday, a 1.6-trillion-parameter multimodal AI model built for long-horizon autonomous coding tasks, trained entirely on domestically produced Chinese AI accelerators for the second time in a row. The model is a mixture-of-experts architecture that activates only approximately 48 billion of its 1.6 trillion total parameters for any given input token, which determines inference cost. Meituan's training cluster consisted of roughly 50,000 domestically produced AI accelerators; the community has identified the likely chip as Huawei's Ascend 910C, though Meituan's announcement uses the phrase "alternative hardware platforms" without naming a vendor. The Ascend 910C delivers approximately 60% of an H100's per-chip performance. Meituan's key architectural innovation is LongCat Sparse Attention, or LSA, which addresses output discontinuity and a quadratic scoring bottleneck in DeepSeek V3's Lightning Indexer through Streaming-aware Indexing, Cross-Layer Indexing, and Hierarchical Indexing. The model also carries 135 billion N-gram Embedding parameters. Pretraining ran across over 35 trillion tokens and reportedly completed without a single rollback or irrecoverable loss spike. The model targets long-horizon autonomous agent tasks, including complex coding jobs.