Kandinsky 3 — open multilingual text-to-image diffusion model from Sber AI. Free weights, local inference friendly.