Mamba-3: Open-Source SSM Beats Transformers on Decode Latency at 1.5B Scale — ICLR 2026 Paper Released by Together AI

Together AI Blog·high signal

Together AI published Mamba-3, a new state space model (SSM) introducing complex-valued state transitions, exponential-trapezoidal discretization, and a MIMO variant — beating Mamba-2, Gated DeltaNet, and even Llama-3.2-1B on prefill+decode latency across all sequence lengths at 1.5B scale. The MIMO variant adds +1.2 points average downstream accuracy on top of the +0.6 point gain from the base architecture vs. the next-best model. Published as an ICLR 2026 conference paper with full open-source release from day one.