DeepHermes 3 - Llama-3 3B Preview

DeepHermes 3 Preview, developed by Nous Research, is a 3.2 billion parameter model built on the Llama-3 architecture, featuring a 32768 token context length. Its primary innovation is the unification of traditional "intuitive" LLM responses with long chain-of-thought reasoning, which can be activated via a specific system prompt. This allows the model to deliberate extensively before providing a solution, improving accuracy for complex problems.

Key Capabilities

• Hybrid Reasoning: Seamlessly switches between intuitive and deep reasoning modes, enabling detailed internal monologues for problem-solving.
• Enhanced Function Calling: Supports structured function calls using a defined system prompt and JSON schema, facilitating integration with external tools and APIs.
• Structured Outputs (JSON Mode): Capable of generating responses strictly adhering to a provided JSON schema, useful for structured data extraction and generation.
• Improved General Performance: Builds upon its predecessor, Hermes 3, with advancements in agentic capabilities, roleplaying, multi-turn conversation, and long context coherence.
• User Steerability: Designed with an emphasis on aligning LLMs to the user, offering powerful steering capabilities through system prompts.

Good For

• Complex Problem Solving: Ideal for tasks requiring deep deliberation and systematic reasoning, where long chains of thought are beneficial.
• Agentic Workflows: Its function calling and structured output capabilities make it suitable for building AI agents that interact with tools and APIs.
• Applications Requiring Control: Developers needing fine-grained control over model behavior and response style through system prompts.
• Benchmarking Reasoning: Offers a unique platform to evaluate reasoning capabilities with and without explicit chain-of-thought activation.