State space models (the Mamba family and successors) are a neural sequence architecture that processes tokens with constant memory per step, avoiding the quadra
A neural sequence architecture (the Mamba family and successors) that processes tokens with constant memory per step, avoiding the quadratic attention costs of transformers on long sequences.
Not wholesale. In 2026 they appear mostly in hybrids that mix SSM layers with attention, targeting long-context and streaming workloads where transformer costs bite hardest. Transformers still dominate frontier models, but SSMs keep architecture competition alive.