PrismML apparently was able to shrink BABA's Qwen 3.6 (27B parameter model) down to less than 4GB (from 54GB) so that it can run locally on an iPhone 17 Pro
$AAPL is in talks with PrismML to use their technology