Chinese models are sometimes better, even if they're distilled (dualuse.dev)

Recent advancements in Chinese AI models have sparked debate about their efficacy and the implications of distillation practices. A White House memorandum from April 23, 2026, criticized these models for allegedly stemming from unauthorized distillation of American counterparts, suggesting they typically underperform. However, the release of GLM 5.1 challenges this narrative, as it has shown superior performance compared to the original Opus 4.6 on several cybersecurity benchmarks, including sophisticated internal evaluations that detect and exploit vulnerabilities. This performance shift indicates that distilled models can indeed surpass their predecessors in specific areas. Additionally, smaller models like Qwen 3.6 Plus have demonstrated impressive capabilities, even outperforming Opus 4.6 in certain cybersecurity scenarios while being more cost-effective. This evolution highlights a significant turning point in the AI landscape, where distillation does not necessarily equate to diminished performance. While leading American models like Opus 4.7 and GPT 5.5 still hold the edge overall, the ability of these Chinese models to excel in targeted applications raises important questions about the trajectory of AI development and the competitive dynamics between different global players in the technology industry.