State Space Models in Production: Mamba and Hybrids

Q: What is a state space model (SSM) in the context of language models?

A state space model is a sequence architecture that processes a sequence by carrying a compressed, fixed-size hidden state through it with a linear recurrence, rather than by comparing every token to every other token the way self-attention does. The lineage runs from classical structured state space models such as S4, which could model long sequences with a linear recurrence computed efficiently as a convolution, through to Mamba, which added a selective mechanism that lets the state-space parameters depend on the input. The practical significance for language modelling is the cost profile: because the state is a fixed size, processing scales linearly in sequence length and autoregressive generation does not accumulate an ever-growing key-value cache, which is the opposite of a Transformer's behaviour. That makes SSMs attractive precisely where attention is most expensive — very long or streaming contexts. The trade is that a fixed-size state, by construction, compresses history, so information genuinely squeezed out of the state is gone, which is a feature for streaming signals and a liability for tasks that need to recall an exact token from far back in the prompt.

Q: How is Mamba different from a Transformer?

The core difference is the mechanism for mixing information across a sequence. A Transformer uses self-attention, which compares every pair of tokens, giving it strong precise recall but a compute cost that grows quadratically with sequence length and a key-value cache that grows with the context during generation. Mamba replaces that with a selective state space model: a linear recurrence that carries a fixed-size state, made powerful by letting the state-space parameters be functions of the input so the model can choose what to propagate or discard as it scans. The Mamba authors report the architecture scales linearly in sequence length and delivers fast inference with roughly five times higher throughput than Transformers, and describe it as a simplified end-to-end network without attention or even MLP blocks. The consequence is that Mamba is far cheaper at long context but weaker at exact in-context recall and copying, because a compressed state cannot match the all-pairs precision of attention. That asymmetry — cheaper propagation, weaker precise recall — is the entire reason the two are combined rather than one simply replacing the other.

Q: Is the Transformer being replaced by state space models?

No, and framing it as replacement is the mistake behind both the hype and the disappointment. What is actually happening in 2026 is hybridisation: production models interleave a majority of cheap, linear-time SSM layers with a minority of attention layers, spending expensive attention only where it earns its cost. The reason is concrete. Pure SSMs have a documented weakness at exact in-context recall, copying and strict format-following — the capabilities Transformer induction heads provide — which disqualifies them for a large class of enterprise work like extraction, structured output and tool-calling. Interleaving a few attention layers restores that behaviour while keeping most of the linear-cost benefit, because only those few layers carry a key-value cache. AI21's Jamba is the canonical example of this hybrid Mamba-Transformer design. So the honest picture is not a successor dethroning an incumbent; it is a portfolio, where the architecture question is how many premium attention layers a given workload truly needs rather than whether attention can be driven to zero. For many short-context, recall-heavy workloads the answer is still a plain Transformer.

Q: What is the biggest weakness of pure state space models?

Exact in-context recall and copying — pulling a specific token, identifier or format out of the prompt verbatim, and following few-shot patterns precisely. This is the capability that Transformer induction heads provide and that a compressed, fixed-size state struggles to match, and it is the single biggest reason pure SSMs have not displaced attention. The most credible source is the Jamba team's own analysis of why they built a hybrid: in their ablations the pure Mamba model often did not follow the correct output format, which they attributed to a limitation of SSMs — a potential difficulty in in-context learning — crediting the Transformer's induction-head mechanism for the capability the SSM lacks. In practical terms, a model that cannot reliably echo an order number, reproduce a schema, or hold to a specified format is disqualifying for extraction, structured output and tool-calling, no matter how cheaply it serves long context. This is why the production architecture is a hybrid: a minority of attention layers restores the recall behaviour while the SSM majority keeps the serving cost low. Where recall genuinely does not matter — long streaming signals, some audio, genomics, time-series — a pure SSM is a legitimate and efficient choice.

Q: What is Jamba and why is it a hybrid?

Jamba is AI21 Labs' hybrid Mamba-Transformer model family, which combines state space model layers with a minority of Transformer attention layers and a mixture-of-experts component, and which the vendor documents at a 256K-token context window. It is a hybrid by deliberate design rather than convenience: the team's published ablations showed that a pure Mamba model often failed to follow the correct format, a recall and in-context-learning weakness they traced to SSMs lacking the Transformer's induction-head mechanism, so they reintroduced attention in a small proportion of layers to restore that capability. The payoff of keeping most layers as SSM is serving economics at long context: the Jamba paper reports that with 128K tokens the model's throughput is three times that of Mixtral on comparable hardware, a concrete demonstration that the SSM component does real economic work rather than winning an abstract benchmark. Jamba matters as the canonical proof that the right way to use state space models in production is compositional — capture the linear-cost benefit across the bulk of the network, and buy back precisely the amount of attention needed to fix the one thing SSMs get badly wrong, no more.

Q: When should I actually use an SSM or hybrid model in production?

Use one when long or streaming context is your dominant serving cost and your workload tolerates the recall profile you can afford — and for most teams that means adopting a released hybrid rather than training a pure SSM yourself. The diagnostic sequence is: first, measure where your serving spend actually goes, because if the key-value cache and quadratic attention at long context are not dominating, an SSM is a solution to a problem you do not have and a well-tuned Transformer with good cache management and speculative decoding is the mature choice. Second, characterise how much of your workload needs exact recall, strict formats or precise in-context lookup, since that number decides how much attention you must keep and whether a pure SSM is ruled out. Third, if you are serving very long documents, logs, audio streams or agent transcripts where the cache is what breaks your budget, stand up a hybrid on your existing stack and benchmark it against your real extraction and format tasks at your actual context lengths. Reserve pure SSMs like Falcon Mamba for the recall-light, long-context corners. The symmetric mistake to avoid is reaching for an SSM because it is new when your workload never stresses attention's cost, or refusing one when long-context economics are exactly the problem it solves.

Q: What is state space duality (Mamba-2)?

State space duality is the theoretical framework introduced in the Mamba-2 paper showing that state space models and attention are not unrelated mechanisms but are formally connected through structured matrices — specifically, that selective SSMs and certain variants of attention can be expressed in a common mathematical form. The paper frames this as theoretical connections between SSMs and variants of attention via structured semiseparable matrices, and uses it to derive a redesigned core layer that runs roughly two to eight times faster than the original Mamba selective scan while remaining competitive with Transformers. The practical importance of this duality is twofold. First, it demystifies why hybrids work so well: if attention and SSMs are points on a shared spectrum, then interleaving them is not bolting together two alien systems but choosing, layer by layer, how much of a shared underlying computation to spend on precise all-pairs recall versus cheap linear propagation. Second, it gives implementers a path to faster kernels by borrowing the well-optimised machinery developed for attention. Duality is therefore both an explanation and an engineering tool, and it is a large part of why the field moved decisively toward composition rather than treating the choice as binary.

Q: Do state space models eliminate the KV cache problem entirely?

A pure state space model does eliminate the growing key-value cache, because it carries a fixed-size recurrent state instead of retaining per-token keys and values, so generation memory does not balloon with context length — which is exactly why SSMs are attractive for very long and streaming contexts. But two important qualifications apply. First, a hybrid model, which is what most production deployments actually use, still has attention layers, so it retains a key-value cache for those layers; the win is that only a small minority of layers carry it, so the cache is far smaller than a full Transformer's rather than absent. Second, a fixed-size state is not free lunch: information that is genuinely compressed out of the state is gone, so the memory saving is paid for in recall precision, which is the SSM weakness that motivates keeping some attention in the first place. So the accurate statement is that SSMs trade an unbounded, precise cache for a bounded, lossy state. For workloads where that trade is favourable — long context, recall-light — it is a large efficiency win; for recall-heavy workloads it is the wrong trade, which is why the production answer is a hybrid that keeps just enough cache to preserve the recall that matters.