I mean, this depends on your sampler. With temp=1 and sampling from the raw output distribution, setting aside numerics issues, these models output nonzero probability of every token at each position