China Narrows LLM Gap to Record-Low 6% of U.S. Performance彭博智库：中国大模型性能差距收窄至历史低点，仅落后美国6%

Bloomberg Intelligence's Thursday report put China's average LLM performance gap versus U.S. models at six percent in June, down from nine percent in May and well below the ten-to-fifteen percent range that prevailed over the prior twelve months. Two Chinese models claimed top-ten positions on LiveBench's global ranking last month.

The compression coincides with a string of high-profile releases — Zhipu's GLM-5.2 topped the global agentic coding leaderboard, and Kimi K3's 2.8-trillion-parameter launch last week rattled assumptions about the ceiling of Chinese model capability.

Bloomberg Intelligence flagged that the trend raises questions about the durability of U.S. technological leadership, though analysts note that inference efficiency and hardware independence remain structural gaps that raw benchmark scores do not fully capture.