Mamba-3
The latest iteration of the Mamba series of state space models. It introduces improvements such as complex-valued states and trapezoidal discretization to approximate continuous-time systems more precisely and to enhance sequence modeling capability.
Timeline (last 90 days)
SteerTTS-edge's acoustic model is a 9.25-million-parameter attention + bidirectional Mamba-3 architecture, compressed to int4 precision through quantization-aware training (QAT).
Each encoder and decoder block in SteerTTS contains, in order, multi-head self-attention, a bidirectional Mamba-3 branch and a convolutional FFN, where the Mamba-3 branch is formed by summing the forward and backward layers, using complex-valued states and trapezoidal discretization.
All Facts (2)
SteerTTS-edge's acoustic model is a 9.25-million-parameter attention + bidirectional Mamba-3 architecture, compressed to int4 precision through quantization-aware training (QAT).
50%UnverifiedEach encoder and decoder block in SteerTTS contains, in order, multi-head self-attention, a bidirectional Mamba-3 branch and a convolutional FFN, where the Mamba-3 branch is formed by summing the forward and backward layers, using complex-valued states and trapezoidal discretization.
50%