March 2026 Summaries

Mamba-3, developed by Cartesia's Chief Scientist Albert Gu, represents the latest advancement in state space models (SSMs) designed for efficient inference, addressing the increasing demand for real-time interaction in modern AI systems. Building on the success of Mamba-2, which prioritized training efficiency, Mamba-3 is optimized for post-training and deployment phases, focusing on inference-heavy tasks. This new model enhances the expressivity of SSM mechanisms through a generalized recurrence, complex-valued state systems, and a multi-input, multi-output (MIMO) framework, significantly improving performance without increasing inference latency. Mamba-3 also incorporates architectural revamps like QKNorm for training stability and the removal of short causal convolutions, which are replaced by more efficient recurrence mechanisms. The model demonstrates superior performance in language modeling tasks compared to its predecessors and other linear alternatives, while also offering notable improvements in retrieval tasks through a hybrid approach that combines linear layers with self-attention. Additionally, Mamba-3 is supported by an open-sourced kernel framework utilizing Triton, TileLang, and CuTe DSL to maximize GPU efficiency, ensuring that the model's performance scales effectively with hardware capabilities.