The "distilled+quantized versions" are not the same model at all, they are existing models (Llama and Qwen) finetuned on outputs from the actual R1 model, and are not really comparable to the real thing.