State-Space Models (Mamba)
Concept
Core Technical
The main architectural challenger: linear-time sequence processing.
You're looking at a concept explainer: what this term means in plain English, and the records it ties together.
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.
How this connects to the rest of the stack
This record's links at a glance — click a group to jump to it.
Basis for
Related
Sources
Record reviewed as of 2026-01-15 — see Methodology.