OpenAI’s Astra has been described across social feeds as running on “recurrent depth” or a “looped transformer,” language that implies a fundamentally new kind of model. Machine learning researcher and educator Sebastian Raschka disputes that framing. His read: the mechanism is real, but it is a narrow engineering choice, not the reason Astra performs well.
Raschka traces the pattern through Nanbeige 4.2-3B, an open-weight model he wrote about roughly two months ago. That model was pretrained on 28 trillion tokens using a design that reuses its 22-layer stack twice rather than once, stretching the effective depth to 44 layers without duplicating a single weight. The parameters stay the same size on disk. The computation the model does at inference time roughly doubles, because every token now travels through nearly twice as many layers before producing an output.
That split matters more than it looks. As inference volume, not training runs, increasingly sets the AI industry’s largest recurring cost, an architecture that trades a smaller footprint for a heavier per-token bill is a deliberate bet on which side of that ledger a lab wants to optimize.
According to Raschka, Nanbeige’s own technical report found that two passes through the layer stack delivered the best tradeoff, keeping roughly three quarters of the per-token efficiency a conventional unlooped design gets. Additional passes barely improved results while making training considerably slower and more expensive, the report found. He notes the underlying idea is not new: it extends a NeurIPS paper on “Mixture-of-Recursions,” which added a learned router deciding, token by token, whether a given piece of text needs one pass through the layers or several.
Raschka’s central judgment is blunt. Astra, he writes, may well be a strong model, but that should not be credited to the looped-layer mechanism, which he calls a small architectural tweak rather than a breakthrough. He also pushes back on a separate claim circulating alongside the Astra coverage: that looped transformers work by hiding some of the model’s chain-of-thought reasoning from view. Layer reuse on its own, he argues, is not what makes visible reasoning text disappear. It adds computation inside hidden states before a token is ever emitted, the same as any ordinary transformer layer does. If a model with more recurrent passes ends up producing fewer visible reasoning tokens, Raschka says the more plausible explanation is that some of its work has simply shifted into latent activations, an effect that would show up just as easily from scaling up model size the conventional way.
For teams evaluating Astra or similar architectures, the useful question is not whether the model loops, but what its inference cost looks like at production volume against the RAM it saves on the way in.
Sebastian Raschka, writing on his own blog on September 2, 2026.
