The five families and what they sell
Every credible alternative falls into one of five families. Knowing the categories first makes the rest of this quest feel less like a parade of acronyms and more like a tour of a small-and-knowable ecosystem.
State Space Models (SSMs / Mamba family). Compress all of history into a fixed-size hidden state. O(n) training (via convolution or selective scan) and O(1) memory at inference. The lineage runs S4 → S4D → H3 → S6/Mamba → Mamba-2 → Mamba-3, with each step softening earlier limitations. Sells: extreme efficiency at long context. Gives up: precise recall over the full window.
RWKV — RNN spirit, Transformer flavor. A linear-attention/RNN hybrid that trains in parallel and runs as an RNN. Names its parameters R/W/K/V, ships at 1.5B Windows devices (the largest deployment of any post-Transformer architecture in the world). Sells: O(1) inference at training-friendly cost. Gives up: long-range exact recall past ~28K tokens, ecosystem maturity.
RetNet — retention as the unifying primitive. Bridges attention and recurrence with fixed exponential decay; supports parallel, recurrent, and chunkwise modes from the same parameters. Has had less production traction than the others but more conceptual influence — it directly seeded Mamba-2's State Space Duality, Google's Hawk/Griffin, and Gated Linear Attention.
Hyena and convolutional alternatives. Implicit long convolutions generated by a small FFN, applied via FFT in O(L log L) time. Found its killer app in genomics (Evo 2, 40B, 1M context) rather than chatbots. Sells: dramatic speedups on very long sequences. Gives up: recall, naturalness on language data.
Linear / efficient attention variants. Performer, Longformer, BigBird, Mistral-style sliding window, NSA, MoBA, Kimi Linear, MHLA. These keep the attention frame and reduce its cost via kernelization, sparsity, or restructuring. Sells: drop-in compatibility with the Transformer ecosystem. Gives up: usually a small quality margin in exchange for usually-large efficiency wins.
And the one that actually ships
Hybrids. Every successful production alternative in 2026 is a hybrid: Jamba (1:7 attention:Mamba), Nemotron-H (~92% attention layers replaced by Mamba-2), Granite 4.0 (9:1 Mamba:attention). They give you most of an SSM's efficiency while keeping enough attention layers for the recall a pure SSM can't do. The narrative arc of this quest is, in one sentence: the field tried to replace attention, learned that pure replacement loses recall, and converged on hybrids.