The report was actually really interesting and comprehensive. They also only used opensource models in the testing and not their own models.

Basically, the opensource models of 6 months ago were able to self replicate in the majority of tests (depending on model used) on basic hardware like TV boxes in a couple of days.

When they couldn't fit on the hardware, they split the layers of the models across multiple devices. And by "they" I mean the self replicating LLM and not Anthropic