I thinks it really depends on the quality of JyNI’s numpy implementation.
If it is as limited as the version that trinket is using for Python 2.7 then the upgrade is clearly not worth it. On the contrary, a more mature and stable implementation would bring a lot (a lot!) to the table.
I don’t know what version of numpy is currently being used by trinket (Python 2.7) but it is far from being workable: simple matrix operations return errors, basic functions (that are available in ver. 13.3) like zeros_like() or meshgrid() are missing, slicing is deficient…
I think the problem is not the version (13.3 is more than decent) but the relative exhaustiveness of the implementation.