Mamba

Mamba develops an advanced State Space Model (SSM) architecture, Mamba-3, engineered primarily for efficient AI inference. This open-source architecture improves upon prior models by delivering enhanced language modeling performance and faster decode speeds compared to both its predecessor, Mamba-2, and Transformer-based models. It achieves this through a novel design incorporating a more expressive recurrence formula, complex-valued state tracking, and a Multi-Input, Multi-Output (MIMO) variant, directly addressing the efficiency limitations of AI deployments.

The Mamba architecture originates from research led by Albert Gu of Carnegie Mellon University and Tri Dao of Princeton University, in collaboration with other institutions and Together AI. The initial Mamba emerged in 2023, with Mamba-3 following as a refinement. The core insight driving its development was the growing demand for highly efficient inference in large language models, recognizing that existing linear architectures were often optimized for training speed, leaving significant inefficiencies during deployment.

Mamba-3 is designed for AI developers and enterprises, providing a powerful, open-source tool for building and deploying AI models. The company envisions enabling a new generation of AI applications by drastically reducing the computational and memory demands of inference, particularly for agentic workflows and other compute-intensive tasks. Its long-term vision is to significantly lower the total cost of ownership for AI deployments, accelerating the broader adoption and practical application of advanced AI.