Preprint finds shared model blocks cut GPU memory use by 18-48%

Researchers report that sharing parts of related language models cut GPU memory use by 18-48% in their tests. Their preprint describes LinkerLLM, a loader that lets variants derived from the same pretrained model share blocks of weights, the values that govern model behavior. The approach fit up to five 7B-parameter variants on one 24 GB consumer GPU. Post-training changed every tensor the researchers examined, so identical-file matching saved no memory. Five of eight tested configurations retained at least 94% of the quality of an unshared variant on the reported benchmarks; each of the other three fell below that threshold on one benchmark. Independently trained specializations showed substantially less shared structure.