State-Space Architecture
The state-space architecture is a blueprint for AI models that read text or signals step by step while carrying along only a compact summary of what has been read so far. It is considered a counter-design to today's common Transformer approach, because it handles very long inputs far more economically.
A language model must somehow retain what came before while reading a text. The models common today simply retain everything: every previous word remains accessible and is reconsidered with every new word. The state-space architecture does it differently. It maintains what is called a state, that is, a fixed, compact summary of everything read so far. With every new word, this state is updated, and the old word itself is no longer needed. The name comes from the fact that the model is always located at a point in a space of possible states and keeps moving through this space.
Why the fixed state saves costs
The advantage becomes apparent with long inputs. In the usual design, the model compares every word with every other word. If you double the length of the text, the computational effort quadruples. With a book of a hundred thousand words, this quickly becomes unaffordable. A state-space architecture, by contrast, grows only linearly: twice as much text, twice as much computing time.
Memory is just as important. Common models store intermediate results for every previous word, which have to sit in graphics card memory. This storage keeps growing with every word. A state-space model always needs the same amount of space, whether it has read ten words or ten thousand. That’s why companies interested in running AI on phones or in cars, where memory is scarce, are paying attention to it.
The price for this is a genuine drawback. Once something has been melted down into the summary, it can no longer be retrieved exactly. A model with full access to all words can quote a phone number from page one verbatim. A state-space model can only do this if it deemed the number important enough to keep in its state.
From state to next word
You can picture the state like a listener’s notes. The listener hears each sentence once and is allowed to write on a sheet of paper of fixed size. With every new sentence, they decide what to add and what to cross out because there’s no more room. In the end, they answer based only on their notes, not on the original.
Mathematically, this is based on equations from control theory, the field that describes things like autopilots or heating controls. There, it has been standard practice for decades to describe a system through a state and a rule for how it changes. AI research has adopted these rules and made the numbers within them learnable. So during training, the model itself learns what it should remember and what it may forget.
Early variants had a fixed memory rule for every input step, which was too rigid. The best-known current variant is called Mamba. It lets the memory rule depend on the input itself. A meaningless filler word then barely changes the state, while an important name changes it strongly. It was precisely this trick that first brought state-space models close to the quality of established language models.
Where these models appear today
In products for end users, pure state-space models are still rarely encountered. They are more common in research models and in hybrid forms, where some layers use the classic design and others use the state trick. Such hybrids are offered by, among others, the AI companies AI21, Nvidia, and Mistral. The reason for mixing is division of labor: one part remembers details word for word, the other carries the long-range context cheaply.
The architecture is especially well suited wherever data occurs as long, uniform streams. Examples include audio recordings, sensor measurements, stock market price series, or DNA sequences with millions of building blocks. With such volumes of data, common models fail due to their costs, while a fixed state simply keeps going.
In news reports, the terms State Space Model, SSM, or Mamba therefore usually appear in connection with two promises: very long inputs and low operating costs. A common misconception is that the architecture is a replacement for Transformers. So far, it is more realistically a complement, and which design will prevail in the long run remains an open question.