Loss of an untrained model

A freshly initialised model with a 32,000-token vocabulary gives roughly equal probability to every token. What cross-entropy loss should you see at step 0? (ln 32,000 ≈ 10.37)