Meituan open-sources 1.6T-parameter LongCat-2.5 agent model trained on Chinese chips

The notable signal is not the benchmark gap but the training stack: a frontier-scale MoE built entirely on Chinese hardware, released under a permissive license with a self-hosting guide and a quantized variant. That lowers the barrier for agent-model deployment outside US compute supply chains, and puts Meituan's research arm in the same conversation as the leading closed labs. Whether the near-parity on Terminal Bench translates into real agentic workloads, and how quickly the Ascend-based recipe is replicated by others, is the open question worth watching.