So it looks like the new ML-kit run-time performance is still pretty bad (although their simple non-tail function-call speed is faster than ours (seen from fib and tak)).