Miami — SubQ claims 56x faster large language model inference than leading models. Dense attention multiplies every token against every other, scaling quadratically; SubQ selects relevant pairs dynami
MIT Technology Review