Mamba 2
The second-generation version of the state space model architecture, which introduces improvements such as a head structure and optimizes training and inference efficiency.
Timeline (last 90 days)
HybridMamba combines Mamba 1's per-coordinate decay, Mamba 2-style head structure, and Mamba 3's complex rotations and dual-endpoint writes, mixing real state blocks with rotated state blocks of different decay laws.
The SISO configurations of Mamba 1-3 can be recovered as exact special cases of HybridMamba.
Retrieval rank refers to the number of independent value channels that a model can write to and read from in parallel at a local interface, and is a key structural dimension distinguishing variants such as Mamba 1, Mamba 2, Mamba 3 and HybridMamba.
On five large-scale H100 bfloat16 workloads, the authors' fused GPU kernels were roughly on par with Mamba 2 within the SSM core scope.
In small-model multi-token recall experiments, both HybridMamba and SSD-MIMO designs outperformed Mamba 1-3 at the hardest level.
On single-seed language models of about 500 million parameters and 2B tokens, HybridMamba and SSD-MIMO outperformed Mamba 1/2 on test loss, but the gains from routing did not translate into lower language model loss.
All Facts (3)
HybridMamba combines Mamba 1's per-coordinate decay, Mamba 2-style head structure, and Mamba 3's complex rotations and dual-endpoint writes, mixing real state blocks with rotated state blocks of different decay laws.
50%UnverifiedThe SISO configurations of Mamba 1-3 can be recovered as exact special cases of HybridMamba.
50%UnverifiedRetrieval rank refers to the number of independent value channels that a model can write to and read from in parallel at a local interface, and is a key structural dimension distinguishing variants such as Mamba 1, Mamba 2, Mamba 3 and HybridMamba.
50%