K/V cache compression and context shortening / summarisation. And yes, I suspected Quants too.