NousResearch/DeepHermes-3-Mistral-24B-Preview

Warm
Public
24B
FP8
32768
Mar 2, 2025
License: apache-2.0
Hugging Face
Overview

DeepHermes 3 - Mistral 24B Preview

DeepHermes 3 Preview, developed by Nous Research, is a 24 billion parameter model that introduces a novel hybrid reasoning approach. It is one of the first LLMs to unify both traditional, intuitive responses and long chain-of-thought reasoning within a single model, activated by a specific system prompt. This allows for deeper consideration of problems and more accurate solutions, with internal monologues enclosed in <think> </think> tags.

Key Capabilities

  • Unified Reasoning Modes: Seamlessly switches between intuitive and deep, systematic reasoning.
  • Enhanced Function Calling: Supports structured function calls with specific system prompts and JSON schema adherence.
  • Improved Annotation & Judgment: Offers advancements in LLM annotation and judgment processes.
  • User Steerability: Designed with an ethos of aligning LLMs to the user, providing powerful control over model behavior.
  • Llama-Chat Format: Utilizes the Llama-Chat format for multi-turn dialogue and system prompt steerability.

Good For

  • Complex Problem Solving: Ideal for tasks requiring extensive deliberation and systematic reasoning.
  • Agentic Applications: Builds upon the Hermes series' focus on advanced agentic capabilities.
  • Structured Output Generation: Capable of generating JSON-formatted responses based on provided schemas.
  • Multi-turn Conversations: Excels in maintaining coherence and context over long dialogues.

This preview model distills early reasoning capabilities and is designed for users seeking advanced control and sophisticated problem-solving from their language models.