TurboMamba, described in a new arXiv preprint, is a language-model architecture designed to improve both speed and accuracy. The core idea is to combine a large global model—a Mamba-3—with a small Transformer that operates on the global model's output. The global model runs only once for each completed block of tokens, rather than at every step, which is intended to reduce computational overhead.
The small Transformer then uses the global context to refine predictions for individual tokens. The authors claim that this setup yields faster inference and faster training compared with existing approaches, while also improving next-token prediction. The abstract does not include experimental details, and the source is a single announcement, so these claims have not been independently verified.
The title mentions lexical memory, but the abstract is truncated and does not describe that component. As a result, the exact role of lexical memory in the architecture remains unclear from this source alone.