[GaggiX]

The paper shows that the speed is comparable to transformer models, faster with smaller with "long" sequence length like 8k.