Mamba-3
Inference-first SSM with a more expressive recurrence, complex-valued state and MIMO updates; +1.8 points over Gated DeltaNet at 1.5B.
Three changes from SSM discretisation give richer recurrence, complex-valued state for state tracking, and a multi-input multi-output form that adds quality without more decode latency. Matches Mamba-2 perplexity with half the state size.
- Date
- Monday, 16 March 2026
- Lab
- CMU / Princeton et al.
- Kind
- paper
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Avg downstream accuracy at 1.5B vs Gated DeltaNet | +0.6 (+1.8 with MIMO) points | authors |
arXiv v1 2026-03-16; ICLR 2026 (listed as oral). Authors are Aakash Lahoti, Kevin Li, Berlin Chen, Caitlin Wang, Aviv Bick, J. Zico Kolter, Tri Dao and Albert Gu. Institution line and weight release were not verified.
Sources
This record was checked against its sources on 6 October 2026. How we check