Mamba-3
Summary
Mamba-3 is a state-space model designed for inference efficiency with three upgrades: a more expressive recurrence from exponential-trapezoidal discretization a complex valued state tracking and multi input multi output SSMs. It achieves faster prefill latency than Mamba-2 and other linear models while decoding latency remains comparable. The kernels are open sourced using Triton TileLang and CuTe DSL and the article covers architecture changes BCNorm RoPE and MIMO along with empirical results.