SubQ 1.1 Card: Linear-scaling sparse attention with 98% retrieval at 12M tokens [pdf]subq.ai2 pointsmitchwainer6 days ago