Hermes 4 expands Nous Research’s line of neutrally-aligned and steerable models with a new group of hybrid reasoners. Like all of our models, these are designed to adhere to the user’s needs and system prompts, rather than to a company’s ethics code. Hermes users will feel an eagerness for roleplaying and creativity in the model. They’ll also notice a lack of lecturing and sycophancy. Put simply, Hermes users will experience a more pleasant, humanistic interaction.
Because they are “reasoning” models, they have the ability to “spend” more tokens during inference to think longer about hard problems. Hybrids can toggle between reasoning and standard responses by including or omitting a <think> tag with a request. By being hybrid, Hermes improves performance across a number of benchmarks and maintains efficiency when thinking isn’t necessary.
The training dataset for Hermes 4 expands over the dataset for Hermes 3 with 50x more data tokens. Much of the data was synthesized using Atropos, and special care was taken to highlight creative content.