Nous Research has released Hermes 4, a new family of open-weight AI models based on Llama 3.1 architecture, designed to advance hybrid reasoning capabilities. Hermes 4 features toggleable hybrid reasoning via tags, allowing the model to switch between fast responses and deeper, step-by-step reasoning. The models come in multiple sizes, including 14 billion, 70 billion, and 405 billion parameters, and are trained on 50 times more data than their predecessor, utilizing a novel DataForge-based synthesis approach with 5 million samples and 19 billion tokens. Hermes 4 achieves high performance on several benchmarks, scoring 96.3 on MATH-500, 87.2 on MMLU, 81.9 on AIME'24, and 61.3 on LiveCodeBench, outperforming other large models such as Cogito 405B and Qwen3 235B in reasoning and coding tasks. The model incorporates a 30,000-token thinking capacity that reduces runaway reasoning by 65-80%, and includes an anti-sycophancy bias to improve response quality. The development emphasizes user alignment, creativity, and flexibility, with built-in authentication features to prepare for integration with AI and crypto technologies. Hermes 4 is fully open source, marking a milestone in open reasoning models and reflecting a brain-inspired hierarchical reasoning system that surpasses traditional large language models like ChatGPT in complex reasoning tasks.