loading
loading
A non-transformer architecture that processes sequences in linear time by carrying a compressed running state instead of attending to everything.
State space models like Mamba ditch attention entirely, instead maintaining a fixed-size hidden state that's updated as they scan the sequence — closer to an RNN but designed to train efficiently in parallel. The win is linear scaling and no growing KV cache, so they handle very long sequences cheaply and stream at constant cost per token. The open question is whether that compressed state can match attention's precise recall over long contexts; in practice many new models hybridize the two.
Plainly
Think of State Space Model (Mamba) as the brain part that guesses or decides. A non-transformer architecture that processes sequences in linear time by carrying a compressed running state instead of attending to everything.
In practice
Use it when model choice, prompts, latency, cost, or quality affect the product result. In practice, define the owner, input, output, and failure mode before you rely on it.