A compact mixture-of-experts model from ai-sage's GigaChat3.1 family, activating only 1.8B parameters out of 10B total — keeping inference lean while drawing on a larger parameter pool. Despite the 'Audio' label in its name, the available data shows it handles text input and output only, so its actual audio-related capabilities remain unclear from the provided information. It runs under an MIT license with open weights, making it freely inspectable and deployable.