DeepHermes 3 - Mistral 24B Preview is a hybrid language model developed by Nous Research, built upon the Mistral-Small-24B base. It is designed to bridge the gap between traditional instruction-tuned models and reasoning-focused systems, having been fine-tuned through distillation from DeepSeek-R1. This preview version is part of the Hermes series, characterized by its high steerability and alignment with user-defined roles.
The model features a dual-mode architecture that allows it to toggle between "intuitive" responses and a dedicated "deep thinking" mode. The non-reasoning (intuitive) mode provides direct and rapid outputs, making it suitable for general-purpose applications such as roleplaying, creative writing, and standard chat. Users can activate the alternative reasoning mode using specific system prompts to trigger internal chains of thought wrapped in <think> tags.
Beyond its conversational capabilities, DeepHermes 3 is optimized for function calling and structured JSON outputs. It supports agentic workflows using a context window of 32,768 tokens and follows the Llama-Chat prompt format. The model's design focuses on maintaining performance in multi-turn dialogues and complex instruction following while providing the flexibility of a hybrid reasoning system.