Codex offers a -spark model that runs on Cerebras. Not quite 17k tok/s, but _very_ fast nonetheless. Worth a look.