The video codec of MiniMax H3 (H3-VisualVAE). It compresses RGB video into a compact form the model predicts, then decodes it back into frames — 16× spatial and 4× temporal compression. FP16 precision: lower memory, quality still good enough for H3's video output.