An article reports looped models with 2.6 billion parameters, trained on more than 7 trillion tokens, achieving performance equivalent to state-of-the-art language models two to three times larger.
Published on October 30, 2025, the article describes scaling looped language models to 2.6 billion parameters and training them on more than 7 trillion tokens. The author reports performance equivalent to that of state-of-the-art language models two to three times larger.
The result is relevant to engineers exploring looped architectures as a way to scale language models. To assess the claim, consult the original article and check how it defines the performance comparison, which evaluations it reports, and under what conditions the results were obtained; those details are not included in the available summary.