Mamba-3

Mamba-3 is a new state space model (SSM) architecture from Together AI designed with inference efficiency as the primary goal, directly motivated by the surge in inference demand from agentic coding tools (explicitly citing Claude Code) and RLVR post-training workloads. Unlike Mamba-2's training-first simplifications that left inference memory-bound, Mamba-3 revisits the SSM transition structure to make GPU computation more compute-bound. The work represents a meaningful shift in SSM design philosophy toward the quality-efficiency frontier that matters most for production AI deployment today.