Softmax is a function that converts a vector of real numbers into a probability distribution that sums to one. It applies the exponential function to each value and divides by the total, so larger values receive larger probabilities.
In language models it turns the final layer's logits into next-token probabilities over the whole vocabulary, and in attention it normalizes query-key scores into weights. Temperature is a value by which logits are divided before softmax, controlling how peaked the distribution is.