Mamba (Transformer Alternative): The Future of LLMs and ChatGPT?

The article discusses the emergence of a non-attention architecture for language modeling, in particular Mamba, which has shown promising results in experimental tests.

Mamba is an example of a state-space model (SSM). But what is a state-space model?

State-Space Models (SSMs)

State-space models (SSMs) are a class of mathematical models used to describe the evolution of a system over time. These models are widely employed in various fields, including control theory, signal processing, economics, and machine learning. State-space models are particularly relevant in the context of language modeling and non-attention architectures, such as the Mamba model mentioned in the provided article.

Here are the key components and concepts related to state-space models:

State Variables (x):

The central concept in a state-space model is the state variable, denoted as “x.” These variables represent the internal state of the system and evolve over time.

State Equation:

The state equation describes how the state variables change over time. It is typically represented as a first-order linear ordinary differential equation (ODE) in continuous time or a first-order difference equation in discrete time.

  • Continuous Time: x’(t) = Ax(t) + Bu(t)
  • Discrete Time: x[k+1] = Ax[k] + Bu[k]

Output Equation:

The output equation relates the observed outputs of the system to its internal state. It is also a linear equation and is often expressed as:

  • Continuous Time: y(t) = Cx(t) + Du(t)
  • Discrete Time: y[k] = Cx[k] + Du[k]

Matrices (A, B, C, D):

  • The matrices A, B, C, and D are parameters of the state-space model.
  • A represents the system dynamics and governs how the state evolves over time.
  • B represents the input matrix, indicating how external inputs affect the state.
  • C defines the output matrix, specifying how the state contributes to the observed outputs.
  • D is the feedforward matrix, accounting for direct transmission of input to output.

Linear Time-Invariant (LTI) Systems:

State-space models are often designed as linear time-invariant systems, meaning that the parameters (A, B, C, D) are constant over time and the system’s behavior is linear.

Continuous and Discrete Time:

State-space models can be formulated in continuous time (using differential equations) or discrete time (using difference equations). The choice depends on the nature of the system and the available data.

State-Space Models in Machine Learning:

In the context of machine learning, state-space models have been used to capture dependencies and temporal relationships in sequential data, such as natural language sequences.

RNNs (recurrent neural networks) can be seen as a kind of nonlinear state-space model. This can be seen readily by looking at the equation for a simple RNN unit:

h(t) = σ(Ah(t-1) + Bx(t))

y(t) = σ(Ch(t))

(in the above, bias terms are removed for simplicity and σ represents a nonlinear activation function)

Recent advancements, as seen in the Mamba model, involve using state-space models to achieve efficient and scalable language modeling without relying on attention mechanisms.

Efficient Scaling with Convolution:

State-space models can be represented in terms of 1D convolutions, which are computationally efficient. This feature contributes to the scalability and efficiency of models like Mamba.

State-space models offer a flexible framework for modeling complex systems with temporal dependencies, and their application in language modeling represents a novel approach to building efficient and effective language models.

How does Mamba extend the vanilla SSM above?

The authors actually start with a continuous-time SSM, and discretize it.

Here, (1a) and (1b) represent a continuous-time SSM. (2a) and (2b) represent a discretized version of the SSM, with new parameters “A bar” and “B bar”. (3a) and (3b) describe how to represent the SSM as a convolution by forming a new matrix “K bar”.

The discretized matrices for A and B are computed as follows:

The above SSM is still LTI (linear time-invariant) since the matrices do not depend on time. The authors then introduce a selection mechanism such that the matrices can depend on the input (x), thereby making the system no longer time-invariant.

The new algorithm is summarized as follows:

Key Components of Mamba:

Data Selection Mechanism:

Mamba incorporates a simple selection mechanism by parameterizing the state-space model (SSM) parameters based on the input text.

This mechanism helps in formulating processing matrices of a recurrent space, B and C, as functions of the input text, adding expressivity at the cost of generality.

Hardware-Aware Algorithm:

Mamba features a hardware-aware algorithm that switches from a convolution to a scan over features, enhancing the model’s efficiency on existing hardware.

This algorithm focuses on storing the latent state efficiently in memory, minimizing the computational bottleneck associated with moving weights.


Mamba combines the recurrence of previous SSMs with the feedforward block style of transformers, creating a novel architecture.

The model introduces a new model block inspired by SSMs and Transformer models, enhancing its expressiveness.

Selective Matrix Parameters:

The data selection mechanism enables Mamba to parameterize the SSM parameters based on input text, allowing matrices to learn which tokens are most important.

This selectivity enhances the model’s ability to capture relevant information from the input sequence.

SRAM Cache:

Mamba uses a sort of cache called SRAM (Static Random-Access Memory) to store core parameters like linearized A, B, and B matrices, optimizing memory usage.

Advantages of Mamba:


The data selection mechanism and selective matrix parameters enhance the expressiveness of Mamba, allowing it to capture important features in the input sequence.

Efficiency in Long-Context Scenarios:

Mamba addresses computational limitations in long-context scenarios, making it suitable for tasks that require processing information over extended sequences.

Hardware Efficiency:

The hardware-aware algorithm and SRAM cache contribute to the efficient utilization of available hardware resources, optimizing the model’s performance.

Inference Speedups:

Custom CUDA kernels in Mamba result in significant inference speedups, improving the model’s efficiency during evaluation.

Performance Comparisons:

Mamba demonstrates competitive performance, as shown by evaluations against benchmark models like Pythia, highlighting its potential in the landscape of language models.


Mamba’s architecture, built on the foundation of state-space models, suggests scalability advantages in terms of potential accuracy and cost of inference for long-context tasks.

Applications and Comparison to Transformers

Mamba serves as a versatile sequence model foundation, demonstrating exceptional performance across various domains, including language, audio, and genomics. In the realm of language modeling, the Mamba-3B model surpasses similarly sized Transformers and competes on par with Transformers that are twice its size, excelling in both pretraining and downstream evaluation tasks.

For more information, read the original paper here.