State-Space Models (Mamba) ☆

State-space models like Mamba process sequences in linear time with constant memory — no quadratic attention, no ballooning KV cache. Pure SSMs trail transformers on recall-heavy tasks, so production systems hybridize (AI21's Jamba, MiniMax's lightning attention). If context keeps growing, the efficiency argument keeps getting stronger — one of the live architecture debates that could reshape hardware demand.