Looped AI model is claimed to run 12 blocks deep on four blocks of memory

IFM AI says its researchers use a loop’s fixed point as a shortcut to scale model depth with compute rather than more memory. Its 1.6B-parameter model runs 12 blocks deep on four blocks of memory. On the same tokens, IFM AI says it beats a four-block Transformer by 17.5% and comes within 2.4% of a 12-block model needing three times the memory. The group links a paper and open-source code.