Please stop blindly trusting and reporting Alibaba's scores.
For example, Qwen3 30b 3b doesn't have anywhere near a legitimate SimpleQA score of 23.6. Gemma 3 27b does have a SimpleQA score of ~10, and it has far more broad knowledge of the domains covered by the SimpleQA test. There's no polite way of saying this. Alibaba flat out cheated, and did so egregiously. Qwen3 30b 3b has less broad English knowledge than Llama 3.2 3b, and its legitimate English SimpleQA score is no higher than 6.
One clear way of seeing this is the 7.2 SimpleQA score of the Qwen3 30b base model, versus 6.1 for the original Qwen3 32 Base. There's no amount of fine-tuning or thinking that can bring the SimpleQA score up by more than a couple points, and that boost is only due to the mitigation of stupid mistakes. Full recall non-multiple choice knowledge is not something you can fine-tune into a base model. The jump to a score of 23.6 was only achieved by explicitly training on SimpleQA test questions.