All terms
Architectures
Mamba
A state space model with input-dependent selection that scales efficiently to long sequences.
Definition
Mamba is a state space model architecture whose key idea is selective state spaces: its state transition and output parameters depend on the input, letting the model choose what to remember or forget like a learned gate. This overcomes earlier state space models that used fixed transitions. Mamba reaches performance competitive with Transformers on language tasks while scaling linearly with sequence length, making it a notable non-Transformer approach to long-context modeling.