Meituan releases a new model with 1.6 trillion parameters, its general-purpose Agent test surpasses Opus-5, with special optimization for domestic chips

What happened: Meituan's LongCat team released LongCat-2.5, a native multimodal model for long-horizon agent tasks with 1.6 trillion total parameters and about 48 billion activated per token. Key numbers: 1.6 trillion total parameters; about 48 billion activated per token; Terminal Bench v4.0 score of 38.5; single-card Why it matters: Meituan says the model leads Claude-Opus-5 and GPT-5.6-Sol on benchmarks including AutomationBench and LongCatClawBench, though it trails Claude-Opus-5 on Terminal Bench v4.0. Training was redesigned around domestic chips with limited per-card memory.

Meituan's LongCat team released LongCat-2.5, a native multimodal model for long-horizon agent tasks with 1.6 trillion total parameters and about 48 billion activated per token.

Meituan says the model leads Claude-Opus-5 and GPT-5.6-Sol on benchmarks including AutomationBench and LongCatClawBench, though it trails Claude-Opus-5 on Terminal Bench v4.0. Training was redesigned around domestic chips with limited per-card memory.