State space models replace the quadratic attention of transformers with a linear-time recurrence carrying a fixed-size state, and as of 2026 they are shipping inside mainstream hybrid checkpoints such as Qwen3-Next and NVIDIA Nemotron rather than displacing transformers outright ref.
Summary
The architecture is still moving. Mamba-3, published at ICLR 2026 by authors at CMU, Princeton, Cartesia AI and Together AI, claims 2× smaller states, enhanced multi-input multi-output decoding aimed explicitly at hardware efficiency, improved scaling laws that narrow the gap with transformers at equal parameter counts, and cleaner integration with attention layers in hybrid stacks 2026 03 Mamba 3 Iclr. That last point matters more than the others: the dominant deployment pattern is not pure SSM but Mamba-plus-attention hybrids, which is how Qwen3-Next and NVIDIA Nemotron-3 handle long context ref.
Evidence base
- Aug 2025: eMamba reports up to 10× speedup and 48.6× lower energy for Mamba on FPGAs and ASICs, and states no prior hardware-acceleration framework was optimised for edge Mamba deployment 2025 08 Emamba Edge Acceleration.
- Early May 2025: the first dedicated Mamba accelerator chiplet (e-chip-V1) taped out in GlobalFoundries 22nm, from an academic lab rather than a startup 2025 06 Wisconsin Mamba Chiplet Gf22.
- March 2026: Mamba-3 published at ICLR 2026 by CMU, Princeton, Cartesia AI and Together AI, claiming 2× smaller states, MIMO decoding for hardware efficiency, and improved scaling that narrows the gap with transformers at equal parameter counts 2026 03 Mamba 3 Iclr.
- June 2026: Qwen3-Next and NVIDIA Nemotron-3 confirmed as shipping Mamba-attention hybrids for efficient long context; mobile NPUs meanwhile exceed 100 tokens/sec decode and ~11,000 tokens/sec prefill on optimised small models ref.
- June 2026: DUET, a disaggregated prefill/decode accelerator for hybrid Mamba-Transformer models, reports 4× faster time to first token, 1.4× higher throughput and 1.5× lower time between tokens versus a B200 GPU on Nemotron-H-56B, Zamba2-7B and Llama3-8B, on the grounds that SSM recurrences map poorly to matmul-centric accelerators.
- Aug 2026: LowRank-SSM finds Mamba input and output projection layers account for over 60% of per-token runtime at sequence lengths of 1,024 and beyond on FPGA, identifying projection rank as an unexplored hardware design variable.
Open questions
- Do Mamba-3’s smaller states and MIMO decoding hold their parameter-efficiency gains at frontier scale, or does the gap with attention re-open above the sizes reported at ICLR 2026?
- Have the core SSM primitives stabilised enough after Mamba-2 to Mamba-3 that a fixed-function accelerator taped out in 2026 is still relevant in 2029?
- Does the hybrid attention-to-SSM layer ratio converge to a stable value across labs, which is what determines whether disaggregated designs like DUET are worth building?
- Can any commercial vendor reproduce eMamba’s ~48× energy advantage on its own silicon against a like-for-like accuracy baseline, rather than citing the academic 22nm ViT comparison?
Assessment drafted 2026-08-31 from up to 16 KB sources using the technology-scorecard framework; scores are a draft read pending review.