US startup launches AI model that reduces compute by 1,000 times

US-based startup Subquadratic has launched its large language model 'SubQ', which reduces compute by nearly 1,000 times compared to standard models. SubQ is the first model built on a fully sub-quadratic sparse attention architecture, which allows it to identify the context that matters and save compute, the firm said. The model outperforms Claude's Opus 4.7 in long context, Subquadratic said.