Mamba

What is Mamba?

Mamba is a sequence model built on the "selective state space model" idea. Its trick is selectivity: when it reads a token, it doesn't cram everything into its state — it decides what to keep and what to drop. Think of it like taking notes in class: the top student only writes down the important parts, so they work faster.

Why does "selective" matter?

The content drives the state
Classic SSMs use a fixed update rule; Mamba lets the input decide what to remember, so memory is sharper.
Cheaper on long text
It skips what's irrelevant, so the state never gets flooded — very long sequences stay manageable.
Linear complexity
Cost grows linearly with length, so hundreds of thousands of tokens are no big deal.

How does it stack up against Transformers?

On efficiency
Mamba is faster and lighter on long sequences, which suits real-time inference and long documents.
On quality
It already trades blows with Transformers on many sequence tasks, and wins some long-sequence ones.
On ecosystem
Transformers have the mature ecosystem; Mamba is catching up fast, and the two are often mixed.

Where can it be used?

Long-document summarization, DNA sequence analysis, speech processing, code generation — anywhere the sequence is long and the response needs to be real-time, Mamba is a strong contender.

Bottom line: Mamba is the top student who knows what to highlight — selective memory keeps it accurate and cheap on very long sequences.