AI glossary · Advanced · Language models · Efficiency & infra
State space model
A state space model (SSM) is an AI model architecture that reads long sequential data, such as text or audio, one piece at a time from the start, continually overwriting a fixed-size summary called the 'state' with what it has seen so far. It is drawing attention as a way to replace or complement the Transformer.
About this term
- Level
- Advanced · Specialist terms that need the inner workings
- Category
- Models & generation
- Added
- 2026-10-08
How it works
An SSM updates its internal 'state' each time it receives an input token (a small piece of text). The state is a compressed summary of everything read so far, and its size never changes. The model looks at the new token and the previous state together, updates the state, and then produces output based on that state. Attention in a Transformer looks back over every previous token for each new token, but an SSM only needs to look at one summary, so the amount of computation does not grow much as the text gets longer. Recent SSMs such as Mamba improved performance by choosing what to keep and what to forget based on the input (selective SSM).
Why it matters
In a Transformer, computation and memory (including the KV cache) grow quickly as the input gets longer, which makes very long documents, audio or genomic data costly to handle. An SSM's cost grows only in proportion to input length, which is an advantage for long data. When generating an answer, it only needs to carry a fixed-size state, so inference is fast and uses little memory. That is why there is strong interest in SSMs for services where long context and efficiency matter, and for on-device AI.
Things to know
Because all information is squeezed into one small state, SSMs are widely considered weaker than attention at finding a specific sentence in a long text and copying it exactly. For this reason, 'hybrid' architectures that mix SSM layers and attention layers are now common. The 'state space' in the name comes from a mathematical framework originally used to describe systems in control engineering, but from a user's point of view it is enough to think of it as 'a sequential model that carries a summary.' The idea resembles an RNN, but the difference is that it is designed so training can run quickly in parallel.
Example
Mamba is a well-known SSM-based language model architecture and has since been tested as a Transformer alternative in a number of studies and open-weight models. On the product side, hybrid LLMs that alternate SSM layers and attention layers have appeared, promoting lower memory use and fast responses even with long documents as their advantages.