Skip to content

2026ResearchIn progress

Post-transformer architectures for foundation models

Representations that stay useful when the task or the data changes.

The dominant approach to foundation models is a Transformer trained on next-token prediction, then scaled in data and parameters. It works well, but it carries structural costs. Attention grows quadratically with sequence length, context is bounded, and inference gets expensive in exactly the regimes where deployment needs it cheap.

Fig. 1 · Cost versus sequence length. Dashed: quadratic attention. Solid: the sub-quadratic regime we aim for.

We work on post-transformer architectures. The hypothesis has two parts: sub-quadratic sequence models can hold Transformer-level quality at scale, and a shared trunk can learn representations that are not bound to next-token prediction. Both parts are unproven. If they hold, three properties follow.

What follows if the hypothesis holds

Longer context

Longer effective context becomes affordable, because compute no longer grows quadratically with sequence length.

Transfer

Representations transfer across tasks and modalities, because the trunk is not shaped by next-token prediction alone.

Cheaper adaptation

A new task is a head on the shared trunk. Serving runs at sub-quadratic cost.

Fig. 2 · One trunk, many heads. A new task is an interface, not a new pretrain from scratch.

Where this sits

The precedent is real. State-space models (S4, Mamba), linear and recurrent attention, and Transformer–SSM hybrids have all shown that sub-quadratic sequence models can be competitive at scale. The open question is whether those results carry into the pretraining of general foundation models rather than isolated benchmark tasks. That is the question we work on.

What we are building

A post-transformer pretraining stack in which the architecture itself is the experiment. We treat the sequence-mixing layer, the memory mechanism, and the interface between trunk and heads as first-class and measurable, because those choices decide quality and transfer.

Fig. 3 · The stack under test. Each layer is a design choice we measure, not a fixed template.

Neognathae and Kestrel

Research that clears that bar ships as domain-specialized models under Neognathae. The neognathae.com site is in development. Kestrel, a text classifier, is available as a service today — write to contact@auxerta.com to get started.

Recent notes

All notes

Cite · Auxerta Labs. Post-transformer architectures for foundation models. Research note v1.0, 2026. auxerta.com/research