Mamba-2
SSM-attention duality architecture, 2-8x faster than Mamba.
About
Mamba-2 is the successor to the Mamba state-space model, introduced by Tri Dao and Albert Gu in their 2024 ICML paper 'Transformers are SSMs.' Its core contribution is structured state space duality (SSD), a mathematical framework showing SSMs and linear attention are two views of the same semiseparable-matrix computation.
The practical payoff: Mamba-2's SSD layer is 2-8x faster to train than Mamba-1 by exploiting tensor cores, while supporting much larger state sizes (N=64-128 vs N=16) and remaining competitive with Transformers on language modeling. Released open under Apache 2.0 via the state-spaces/mamba repository, it underpins hybrids like Jamba and Falcon-Mamba.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 6/10
Scored 6/10 (metrics tier, metrics-v1). Evidence: Verified via vendor site/docs/GitHub and reputable coverage, wave 66; Trending score 60/100 from researcher assessment.