This isn't really news, but just putting some numbers on why relying on passing data in kwargs doesn't generally scale well: there ends up being a lot of implicit copying to avoid accidental mutation of unowned objects. In general, if speed or memory usage matters and you potentially have a lot of entries in a map, avoid using kwargs as the sole…

cpython_kwargs.sh

# Using even one additional keyword arg forces CPython to make an additional copy of the entire original keyword dictionary
$ python3 -m timeit -s "kwds = {str(k):v for k, v in enumerate(range(1000))}" -s "def f(**kwargs): return kwargs" "f(**kwds)"
10000 loops, best of 3: 50.5 usec per loop
$ python3 -m timeit -s "kwds = {str(k):v for k, v in enumerate(range(1000))}" -s "def f(**kwargs): return kwargs" "f(other=1, **kwds)"
10000 loops, best of 3: 82.4 usec per loop


# Using a non-dict subclass forces CPython to copy it to a real dictionary as part of the call
$ python3 -m timeit -s "import collections; kwds = collections.UserDict((str(k), v) for k, v in enumerate(range(1000)))" -s "def f(**kwargs): return kwargs" "f(**kwds)"
1000 loops, best of 3: 573 usec per loop
$ python3 -m timeit -s "import collections; kwds = collections.UserDict((str(k), v) for k, v in enumerate(range(1000)))" -s "def f(**kwargs): return kwargs" "f(other=1, **kwds)"
1000 loops, best of 3: 612 usec per loop