Compressing giant AI models just got cheaper — a faster math trick replaces the old method

Researchers found a way to compress large pretrained AI models using a mathematical technique (randomized subspace iteration) that works better than the previous standard approach when models have certain structural properties. This matters because it lets companies run big AI models on smaller computers without losing accuracy, which makes deployment cheaper and faster.

Every major AI model today is huge — too big to run on normal servers without eating enormous computational costs. Until now, the standard compression method (randomized SVD) worked poorly on modern models because of how their mathematical structure behaves. This paper shows a different compression algorithm that actually works on these models at scale. The practical upshot is simpler: if this technique gets adopted, the cost to deploy AI models shrinks, which lowers the barrier for companies that can't afford massive inference budgets.