AnalysisAI ModelsSeptember 2, 2026

Raschka debunks Astra 'looped transformer' hype

Sebastian Raschka explains that OpenAI's Astra is not a novel 'recurrent depth' model, but likely uses a looped transformer that reuses layers. He cites Nanbeige 4.2, which reuses a 22-layer stack twice, doubling effective depth without extra parameters, retaining ~75% token efficiency.

2 sources

More stories today

Open the live feed