Hermes 3
About
The Hermes 3 family of large language models (LLMs), developed by NousResearch, represents a significant advancement in generalist instruction models 146. Built upon the Llama 3.1 foundation model, Hermes 3 models are available in 8B, 70B, and 405B parameter versions 146. A key design principle is enhanced steerability, achieved through targeted training to precisely follow system and instruction prompts in a neutral and adaptive manner 146. This leads to models that are highly responsive to system prompts, allowing fine-grained control over behavior and persona 8. In addition to instruction following, Hermes 3 features long-term context retention, multi-turn conversation, complex role-playing, internal monologue abilities, and enhanced agentic function-calling 146. These models excel in structured output generation, utilizing XML tags and scratchpads for transparency and accuracy 1. The training data is a carefully curated blend of approximately 390 million tokens, including a significant portion of synthetically generated responses to encourage precise instruction following and nuanced reasoning 13. While NousResearch claims superior performance over Llama 3.1 in certain areas 46, independent evaluations indicate mixed results, underscoring the challenges in benchmarking and comparing LLMs 3.