Nous Research

Post-training and tools for training

Nous Research develops AI models and training systems, including synthetic-data generation, reinforcement learning and agent software. Its Hermes family adapts existing foundation models through post-training. That is a development contribution beyond serving an unchanged checkpoint, while retaining a dependency on the original model family. [1][2]

It released Hermes 4 on August 25, 2025, initially including 70-billion and 405-billion-parameter models based on Llama 3.1. The technical report describes a training mixture of reasoning and non-reasoning examples, and a hybrid mode that can produce deliberate reasoning or more direct responses. [2][3][4]

Public models with inherited license terms

The Hermes 4 model cards provide downloadable weights and local inference instructions. Their license metadata identifies Llama terms, so public availability should not be described as Apache 2.0 licensing. The cards also describe structured outputs and tool use. These are the developer's documented behaviors and evaluations, not a guarantee that every generated tool call or reasoning trace is correct. [3][4]