NousResearch/Hermes-3-Llama-3.2-3B

Overview

Hermes 3 3B is Nous Research's first fine-tune in the 3 billion parameter class, built upon the Llama-3.2 foundation model. It represents an evolution of the Hermes series, focusing on user alignment and providing powerful steering and control capabilities. The model demonstrates significant improvements across various generalist LLM tasks, including advanced agentic capabilities, enhanced roleplaying, reasoning, and multi-turn conversation coherence, particularly over long contexts. It also features more robust function calling and structured output generation compared to its predecessors.

Key Capabilities

• User Alignment & Control: Designed for strong user alignment with powerful steering capabilities.
• Generalist Assistant: Improved performance in general assistant tasks.
• Advanced Agentic Functions: Enhanced capabilities for agent-like behaviors.
• Improved Dialogue: Better roleplaying, reasoning, and multi-turn conversation coherence.
• Long Context Handling: Maintains coherence over extended conversational contexts.
• Reliable Function Calling: Offers more powerful and reliable function calling mechanisms.
• Structured Output: Enhanced capabilities for generating structured outputs, including JSON mode.

Benchmarks

Hermes 3 3B is competitive with Llama-3.1 Instruct models in general capabilities. It achieves an average of 64.00 on GPT4All benchmarks, 34.36 on AGIEval, and 43.76 on BigBench, showcasing its performance across various reasoning and knowledge-based tasks.

Prompt Format

The model utilizes the ChatML prompt format, enabling structured multi-turn dialogue and system instruction steerability. It supports OpenAI endpoint compatibility and provides specific formats for both function calling and JSON mode/structured outputs, allowing for precise control over model responses.