Hermes 3 sells steerability over guardrails

Hermes 3 matters less for the spooky screenshot and more for the product thesis behind it: open models can compete by being more controllable, more personalizable, and less locked into one lab’s safety style.

• –Hermes 3 is built by fine-tuning Llama 3.1 8B, 70B, and 405B, so the story is not just “new weights” but how much post-training still changes model behavior
• –Nous is explicitly marketing steerability as a feature, which pushes against the closed-model trend of rigid refusals and heavy baked-in defaults
• –The blank-prompt “existential crisis” demo is a vivid proof that assistant personality is a training choice, not an inevitable property of LLMs
• –Outside commentary around the release focused on whether Hermes 3 truly qualifies as a frontier-class model, but even skeptics treated it as a meaningful open fine-tune worth benchmarking
• –For developers, the interesting bit is practical: if Hermes follows system prompts more faithfully, it could be more useful for agents, tooling, and specialized chat workflows than safer but more stubborn closed models