Hermes 3 is NousResearch's fine-tune of Meta's Llama 3.1 70B, tuned for strong instruction-following, agentic behavior, and structured output generation. It handles long contexts well — up to 128K tokens — and tends to follow complex, multi-step instructions more reliably than the base model. The trade-off is that, like most community fine-tunes, its behavior can vary depending on how prompts are structured.