Meituan has released LongCat-2.5, a 1.6-trillion-parameter AI agent model, with its weights now publicly available. This model was trained exclusively on Chinese hardware, specifically 194 Huawei Ascend chips, and notably did not use any Nvidia GPUs. On the Terminal Bench benchmark, LongCat-2.5 achieved a score of 88.4, placing it close to Claude Opus 5. AI
IMPACT Demonstrates capability for large-scale AI model training outside of traditional Nvidia GPU ecosystems.
RANK_REASON Frontier-lab model release with system card. [lever_c_demoted from frontier_release: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →