DeepHermes 3 preview (24B and 3B) from Nous Research unifies reasoning and normal LLM responses into | Learn large language models free | Top llm models huggingface | Llm training dataset free | Turtles AI

DeepHermes 3 Preview represents the most advanced evolution of Nous Research’s Hermes series, introducing a next-generation linguistic model capable of unifying intuitive responses and complex reasoning in a single structure. This pioneering model aims to redefine the interaction with LLMs, improving accuracy, judgment and adaptability through an advanced prompt system.

Key Points:

• Unification of Response Modes: DeepHermes 3 integrates intuitive responses and complex thought chains.
• Advanced Prompt System: Introduces the Llama-Chat format for a more structured management of multi-turn dialogues.
• Hybrid Reasoning Capabilities: Supports both quick responses and in-depth reflections triggered by specific system prompts.
• Adaptability and User Control: Allows the user to customize the interaction with the model, ensuring greater control over the results.

DeepHermes has today unveiled a pair of cutting-edge hybrid reasoning models that promise to reshape how AI handles complex tasks on the fly. Launched on March 13, 2025, the new lineup features DeepHermes 24B and DeepHermes 3B—each engineered to balance robust problem-solving with efficiency.

The DeepHermes 24B model is designed for demanding computational tasks. It harnesses advanced reasoning capabilities to deliver detailed, context-rich outputs ideal for intensive applications. Meanwhile, its streamlined sibling, DeepHermes 3B, is optimized for on-device performance. With its 8‑bit quantization, this variant offers remarkable efficiency, making it a practical choice for devices with modest hardware without sacrificing output quality.

A standout feature across both models is a dedicated switch that toggles advanced reasoning functions. When activated, the models dynamically enhance their analytical performance, providing solutions that go beyond standard language processing. Deactivating this switch reverts the models to their baseline operations, allowing users to choose between power and speed as their needs evolve.

Despite sharing identical core weights, the behavior of these models shifts dramatically based on the reasoning setting, offering a versatile approach that adapts to various use cases. Early evaluations suggest that this dual-mode operation has caught the attention of benchmark evaluators, positioning DeepHermes as a strong competitor against larger, cloud-dependent systems.

With DeepHermes 3, Nous Research opens new perspectives in the field of AI, confirming the importance of adaptable, precise models capable of tackling the challenges of advanced reasoning.