Vocos is an open-weights neural vocoder developed by GemeloAI that generates audio waveforms by producing Short-Time Fourier Transform spectral coefficients rather than directly synthesizing time-domain signals. The model uses a ConvNeXt-based architecture and supports both mel-spectrogram and neural audio codec token inputs, achieving high reconstruction fidelity with computational efficiency up to 70 times faster than comparable vocoders while maintaining perceptual quality scores competitive with leading adversarial models.