**Short answer**: Generally, you don't need to do `softmax` if you don't need probabilities. And using raw logits leads to more numerically stable code. **Long answer**: First of all, the inputs of the `softmax` layer are called [logits][1]. *During evaluation*, if you are only interested in the highest-probability class, then you can do `argmax(vec)` on the logits. If you want probability distribution over classes, then you'll need to exponentiate and normalize to 1 - that's what `softmax` does. *During training*, you need to have loss function to optimize. Your training data contains true classes, so you have your target probability distribution $p_i$, which is 1 at your true class and 0 at all other classes. And you train your network to produce a probability distribution $q_i$ as close to the target as possible. The "distance" measure between two probability distribution is called [cross-entropy][2]: $$ H = - \sum p_i \log q_i $$ As you can see, you also only need logs of the output probabilities - so the logits will suffice to compute the loss. As an example, the `keras` standard [`CategoricalCrossentropy` loss][3] can be configured to compute it `from_logits` and it mentions that: > Using from_logits=True is more numerically stable. [1]: https://en.wikipedia.org/wiki/Multinomial_logistic_regression [2]: https://en.wikipedia.org/wiki/Cross_entropy [3]: https://www.tensorflow.org/api_docs/python/tf/keras/losses/CategoricalCrossentropy