ModelsHigh importance

Meituan launches LongCat-2.5, a 1.6-trillion-parameter Chinese model built to run on domestic chips

Meituan, the Chinese delivery giant, launched LongCat-2.5, a mixture-of-experts (MoE) model with 1.6 trillion parameters, about 48 billion active per token, and multimodal from training. The company says it runs with 4-bit weights and activations (W4A4) on Chinese accelerators with less than 80 GB of memory. In Meituan's own numbers, it beats Claude Opus 5 on BrowseComp (91.9 vs. 90.8) but trails on Terminal Bench 4.0 (38.5 vs. 53.9) and on an internal app-building test (73.7 vs. 76.4). The weights have not been published yet.