Nous Research Team Releases Hermes 4: A Family of Open-Weight AI Models with Hybrid Reasoning

Nous Research has unveiled Hermes 4, a family of open-weight language models spanning 14B, 70B, and 405B parameters, built on Llama 3.1 checkpoints, which achieve state-of-the-art performance solely through post-training techniques. A key innovation is the introduction of hybrid reasoning capabilities, allowing models to switch between standard responses and explicit, step-by-step reasoning using embedded <think> tags, thereby enhancing their ability to handle complex problems with transparency and neutrality. Central to Hermes 4's development is DataForge, a novel graph-based synthetic data generation system that constructs training