Establishing a high-accuracy and training-free brain-computer interface (BCI) system is essential for improving BCI practicality. In this study, we propose for the first time a training-free double-partial least-squares (D-PLS) model for steady-state visual evoked potential (SSVEP) detection that consists of double-layer PLS, a PLS spatial filter, and a PLS feature extractor. Electroencephalographic data from 11 healthy volunteers under four different visual stimulation frequencies were used to test the proposed method. Compared with commonly used spatial filters, minimum energy combination and average maximum contrast combination, the classification accuracies could be improved 2-10% by our proposed PLS spatial filter. Furthermore, our proposed PLS feature extractor achieved better performance than current feature extraction methods, namely power spectral density analysis, canonical correlation analysis, and the use of the least absolute shrinkage and selection operator. The average classification accuracy for our proposed D-PLS model exceeded 90% when the signal time window was longer than 3.5 s and reached as high as 93.9% when the time window was 5 s. Moreover, the DPLS model can be easily set without training data, so it can be used widely in SSVEP-based BCI systems.