Down to what $k$ and how can we devise a practical, public, efficiently computable One Way Permutation $P$ of the set $\{0,1\}^k$ of $k$-bit bitstrings, if possible without involving a trusted party for parameter setup ? Note: I want $P$ to be a permutation in the mathematical sense, and computationally hard to invert for some definition of that.
For values of $k$ starting at bout $2000$, we can use the assumed hardness of the Discrete Logarithm Problem (RSA also works if we trust a party to generate and publish an appropriate public key, and destroy any clue that could lead to factorization of the modulus).
For example, with the DLP, we find the smallest prime $p$ at least $\sqrt[3]{9}\,2^k$ such that $q=(p-1)/2$ is also prime, find the smallest $g$ at least $p/\pi$ such that $g^q\bmod p=1$, and define function $F$ over the set $\{1,2\dots q\}$ as $F(x)=\min\big(g^x\bmod p,p-(g^x\bmod p)\big)$. As far as we know, $F$ is an OWP over that set. We then use cycling to reduce $F$ to a permutation $P$ of $\{0,1\}^k$: we convert the input bitsring to integer, add one, apply $F$ and iterate (on average $\sqrt[3]{9}/2\approx1.04$ times) until the result is at most $2^k$, subtract one, convert back to bitstring.
Addition: I'll cowardly let the answer state its security claim. Informally, I'd be happy with a vague argument that more than $\min(2^n,2^{128})$ cycles of classical CPU are required for anything that should require more than $2^n$ evaluations in the forward direction, like finding a bitstring $x$ with the first $n$ bits of $P(x)$ all-zero (or other arbitrary $n$-bit value defined independently of the definition of $P$).
Motivation: I read that OWPs are more useful than OWFs. I wonder if that matters in practice. I reason that if we can't get $k$ down to say 256, we could as well use as a practical equivalent of a OWP a hash of $k$ bit (and same input size); it's most likely not a OWP, but it can't be computationally distinguished from that if the hash is secure.