How many models should be generated by rosetta for de novo protein structure prediction of a 382 long sequence? In the tutorial, it says 50,000 models but practically it takes a lot of time (weeks) to generate this number. For instance, I am currently running rosetta on a 382 long sequence and in a week it has generated 1500 models; I am using 16 cpus with Openmpi.
Unfortunately you do in general increase your chance of success by going to the 50k range of models or higher. Projects like Rosetta@home exist because even large research groups can't access as much compute power as they'd like. In this case the long time arises from the combinatorial explosion of possible fragment combinations that need to be explored.
Success is rather protein dependent though - sometimes the right structure falls out after only a few models, other times your fragment library does not cover the protein adequately and you could run Rosetta forever without getting a good structure. Have a look at your 1,500 models and see how similar they look.
382 residues is rather long for Rosetta. Another factor to consider is whether you are using predicted contacts as part of the folding process (e.g. in the PconsFold protocol), which will greatly improve the chances of getting a good model and reduce the number of models you need. In fact the PconsFold group switched to CONFOLD for PconsFold2, I think mainly because it was much faster, so you are not alone in finding Rosetta runs long.