Meet Mamba-3: A New State Space Model Frontier with 2x Smaller States and Enhanced MIMO Decoding Hardware Efficiency

Mamba-3 is a new inference-first state space model introduced by researchers from CMU, Princeton, Together AI, and Cartesia AI.

It adds exponential-trapezoidal discretization, complex-valued state updates with a RoPE equivalence, and a MIMO decoding formulation.

MIMO raises decoding arithmetic intensity up to 4x and lets Mamba-3 match Mamba-2 perplexity with half the state size.

On FineWeb-Edu at 1.5B parameters, Mamba-3 MIMO achieved higher downstream accuracy than Mamba-3 SISO and prior models.