“The Mamba architecture offers an alternative to transformers by compressing sequence history into a small state vector, which is particularly beneficial for large-batch inference workloads.”